	.version 2.2
	.target sm_20
	// compiled with ../../../External/3rdParty/NVIDIA/CUDA/win/bin/../open64/lib//be.exe
	// nvopencc 3.2 built on 2010-11-04

	.visible .func (.param .s32 __cudaretf__Z15IntegerMultiplyii) _Z15IntegerMultiplyii (.param .s32 __cudaparmf1__Z15IntegerMultiplyii, .param .s32 __cudaparmf2__Z15IntegerMultiplyii)

	.visible .func (.param .s32 __cudaretf__Z17Standard2DKernelXv) _Z17Standard2DKernelXv ()

	.visible .func (.param .s32 __cudaretf__Z17Standard2DKernelYv) _Z17Standard2DKernelYv ()

	.visible .func (.param .align 16 .b8 __cudaretf__Z13Half4ToFloat47ushort4[16]) _Z13Half4ToFloat47ushort4 (.param .align 8 .b8 __cudaparmf1__Z13Half4ToFloat47ushort4[8])

	.visible .func (.param .align 8 .b8 __cudaretf__Z13Float4ToHalf46float4[8]) _Z13Float4ToHalf46float4 (.param .align 16 .b8 __cudaparmf1__Z13Float4ToHalf46float4[16])

	.visible .func (.param .u32 __cudaretf__Z4Mix3RjS_S_) _Z4Mix3RjS_S_ (.param .u64 __cudaparmf1__Z4Mix3RjS_S_, .param .u64 __cudaparmf2__Z4Mix3RjS_S_, .param .u64 __cudaparmf3__Z4Mix3RjS_S_)

	.visible .func (.param .s32 __cudaretf__Z4Randj) _Z4Randj (.param .u32 __cudaparmf1__Z4Randj)

	.visible .func (.param .s32 __cudaretf__Z6Rand2Djjj) _Z6Rand2Djjj (.param .u32 __cudaparmf1__Z6Rand2Djjj, .param .u32 __cudaparmf2__Z6Rand2Djjj, .param .u32 __cudaparmf3__Z6Rand2Djjj)

	.visible .func (.param .s32 __cudaretf__Z6Rand2Dj) _Z6Rand2Dj (.param .u32 __cudaparmf1__Z6Rand2Dj)

	.visible .func (.param .s32 __cudaretf__Z19MaxUnsignedBitValuei) _Z19MaxUnsignedBitValuei (.param .s32 __cudaparmf1__Z19MaxUnsignedBitValuei)

	.visible .func (.param .f32 __cudaretf__Z13MaxDepthValue14IR_PixelFormat) _Z13MaxDepthValue14IR_PixelFormat (.param .s32 __cudaparmf1__Z13MaxDepthValue14IR_PixelFormat)

	.visible .func (.param .f32 __cudaretf__Z15DepthScaleValue14IR_PixelFormatS_) _Z15DepthScaleValue14IR_PixelFormatS_ (.param .s32 __cudaparmf1__Z15DepthScaleValue14IR_PixelFormatS_, .param .s32 __cudaparmf2__Z15DepthScaleValue14IR_PixelFormatS_)

	.visible .func (.param .s32 __cudaretf__Z7IsYCbCr14IR_PixelFormat) _Z7IsYCbCr14IR_PixelFormat (.param .s32 __cudaparmf1__Z7IsYCbCr14IR_PixelFormat)

	.visible .func (.param .f32 __cudaretf__Z11YCbCrOffseti14IR_PixelFormat) _Z11YCbCrOffseti14IR_PixelFormat (.param .s32 __cudaparmf1__Z11YCbCrOffseti14IR_PixelFormat, .param .s32 __cudaparmf2__Z11YCbCrOffseti14IR_PixelFormat)

	.visible .func (.param .align 16 .b8 __cudaretf__Z15AddAYCbCrOffset6float414IR_PixelFormat[16]) _Z15AddAYCbCrOffset6float414IR_PixelFormat (.param .align 16 .b8 __cudaparmf1__Z15AddAYCbCrOffset6float414IR_PixelFormat[16], .param .s32 __cudaparmf2__Z15AddAYCbCrOffset6float414IR_PixelFormat)

	.visible .func (.param .align 16 .b8 __cudaretf__Z20SubtractAYCbCrOffset6float414IR_PixelFormat[16]) _Z20SubtractAYCbCrOffset6float414IR_PixelFormat (.param .align 16 .b8 __cudaparmf1__Z20SubtractAYCbCrOffset6float414IR_PixelFormat[16], .param .s32 __cudaparmf2__Z20SubtractAYCbCrOffset6float414IR_PixelFormat)

	.visible .func (.param .u64 __cudaretf__Z23ColorSpaceConvertMatrix14IR_PixelFormatS_) _Z23ColorSpaceConvertMatrix14IR_PixelFormatS_ (.param .s32 __cudaparmf1__Z23ColorSpaceConvertMatrix14IR_PixelFormatS_, .param .s32 __cudaparmf2__Z23ColorSpaceConvertMatrix14IR_PixelFormatS_)

	.visible .func (.param .s32 __cudaretf__Z13IsGammaLinear14IR_PixelFormat) _Z13IsGammaLinear14IR_PixelFormat (.param .s32 __cudaparmf1__Z13IsGammaLinear14IR_PixelFormat)

	.visible .func (.param .f32 __cudaretf__Z18ApplyGammaFunctionff) _Z18ApplyGammaFunctionff (.param .f32 __cudaparmf1__Z18ApplyGammaFunctionff, .param .f32 __cudaparmf2__Z18ApplyGammaFunctionff)

	.visible .func (.param .align 16 .b8 __cudaretf__Z16OpaqueComponents6float414IR_PixelFormat[16]) _Z16OpaqueComponents6float414IR_PixelFormat (.param .align 16 .b8 __cudaparmf1__Z16OpaqueComponents6float414IR_PixelFormat[16], .param .s32 __cudaparmf2__Z16OpaqueComponents6float414IR_PixelFormat)

	.visible .func (.param .align 16 .b8 __cudaretf__Z21PremultiplyComponents6float414IR_PixelFormat[16]) _Z21PremultiplyComponents6float414IR_PixelFormat (.param .align 16 .b8 __cudaparmf1__Z21PremultiplyComponents6float414IR_PixelFormat[16], .param .s32 __cudaparmf2__Z21PremultiplyComponents6float414IR_PixelFormat)

	.visible .func (.param .align 16 .b8 __cudaretf__Z23UnpremultiplyComponents6float414IR_PixelFormat[16]) _Z23UnpremultiplyComponents6float414IR_PixelFormat (.param .align 16 .b8 __cudaparmf1__Z23UnpremultiplyComponents6float414IR_PixelFormat[16], .param .s32 __cudaparmf2__Z23UnpremultiplyComponents6float414IR_PixelFormat)

	.visible .func (.param .align 16 .b8 __cudaretf__Z23ConvertPixel_444_To_4446float414IR_PixelFormatS0_[16]) _Z23ConvertPixel_444_To_4446float414IR_PixelFormatS0_ (.param .align 16 .b8 __cudaparmf1__Z23ConvertPixel_444_To_4446float414IR_PixelFormatS0_[16], .param .s32 __cudaparmf2__Z23ConvertPixel_444_To_4446float414IR_PixelFormatS0_, .param .s32 __cudaparmf3__Z23ConvertPixel_444_To_4446float414IR_PixelFormatS0_)

	.visible .func (.param .align 16 .b8 __cudaretf__Z18SwapComponentOrderI6float4ET_RKS1_[16]) _Z18SwapComponentOrderI6float4ET_RKS1_ (.param .u64 __cudaparmf1__Z18SwapComponentOrderI6float4ET_RKS1_)

	.visible .func (.param .align 16 .b8 __cudaretf__Z15ClampComponentsI6float4ET_RKS1_ff[16]) _Z15ClampComponentsI6float4ET_RKS1_ff (.param .u64 __cudaparmf1__Z15ClampComponentsI6float4ET_RKS1_ff, .param .f32 __cudaparmf2__Z15ClampComponentsI6float4ET_RKS1_ff, .param .f32 __cudaparmf3__Z15ClampComponentsI6float4ET_RKS1_ff)

	.visible .func (.param .align 16 .b8 __cudaretf__ZplI6float4ET_RKS1_f[16]) _ZplI6float4ET_RKS1_f (.param .u64 __cudaparmf1__ZplI6float4ET_RKS1_f, .param .f32 __cudaparmf2__ZplI6float4ET_RKS1_f)

	.visible .func (.param .u64 __cudaretf__ZpLI6float4ERT_S2_f) _ZpLI6float4ERT_S2_f (.param .u64 __cudaparmf1__ZpLI6float4ERT_S2_f, .param .f32 __cudaparmf2__ZpLI6float4ERT_S2_f)

	.visible .func _Z7Write2DI7ushort4EvT_PS1_iii (.param .align 8 .b8 __cudaparmf1__Z7Write2DI7ushort4EvT_PS1_iii[8], .param .u64 __cudaparmf2__Z7Write2DI7ushort4EvT_PS1_iii, .param .s32 __cudaparmf3__Z7Write2DI7ushort4EvT_PS1_iii, .param .s32 __cudaparmf4__Z7Write2DI7ushort4EvT_PS1_iii, .param .s32 __cudaparmf5__Z7Write2DI7ushort4EvT_PS1_iii)

	.visible .func _Z7Write2DI6float4EvT_PS1_iii (.param .align 16 .b8 __cudaparmf1__Z7Write2DI6float4EvT_PS1_iii[16], .param .u64 __cudaparmf2__Z7Write2DI6float4EvT_PS1_iii, .param .s32 __cudaparmf3__Z7Write2DI6float4EvT_PS1_iii, .param .s32 __cudaparmf4__Z7Write2DI6float4EvT_PS1_iii, .param .s32 __cudaparmf5__Z7Write2DI6float4EvT_PS1_iii)

	.visible .func (.param .align 16 .b8 __cudaretf__Z18UnpremultiplyPixel8PixelRGB[16]) _Z18UnpremultiplyPixel8PixelRGB (.param .align 16 .b8 __cudaparmf1__Z18UnpremultiplyPixel8PixelRGB[16])

	.visible .func (.param .f32 __cudaretf__Z13ToLinearColorf) _Z13ToLinearColorf (.param .f32 __cudaparmf1__Z13ToLinearColorf)

	.visible .func (.param .f32 __cudaretf__Z15FromLinearColorf) _Z15FromLinearColorf (.param .f32 __cudaparmf1__Z15FromLinearColorf)

	.visible .func (.param .align 16 .b8 __cudaretf__Z25PremultiplyLinearizePixel8PixelRGB[16]) _Z25PremultiplyLinearizePixel8PixelRGB (.param .align 16 .b8 __cudaparmf1__Z25PremultiplyLinearizePixel8PixelRGB[16])

	.visible .func (.param .align 16 .b8 __cudaretf__Z29UnpremultiplyUnlinearizePixel8PixelRGB[16]) _Z29UnpremultiplyUnlinearizePixel8PixelRGB (.param .align 16 .b8 __cudaparmf1__Z29UnpremultiplyUnlinearizePixel8PixelRGB[16])

	.visible .func (.param .align 16 .b8 __cudaretf__Z20PremultiplyLinearize6float4[16]) _Z20PremultiplyLinearize6float4 (.param .align 16 .b8 __cudaparmf1__Z20PremultiplyLinearize6float4[16])

	.visible .func (.param .align 16 .b8 __cudaretf__Z24UnpremultiplyUnlinearize6float4[16]) _Z24UnpremultiplyUnlinearize6float4 (.param .align 16 .b8 __cudaparmf1__Z24UnpremultiplyUnlinearize6float4[16])

	.visible .func (.param .align 16 .b8 __cudaretf__Z23ConvertPixel_422_To_4446float414IR_PixelFormatS0_[16]) _Z23ConvertPixel_422_To_4446float414IR_PixelFormatS0_ (.param .align 16 .b8 __cudaparmf1__Z23ConvertPixel_422_To_4446float414IR_PixelFormatS0_[16], .param .s32 __cudaparmf2__Z23ConvertPixel_422_To_4446float414IR_PixelFormatS0_, .param .s32 __cudaparmf3__Z23ConvertPixel_422_To_4446float414IR_PixelFormatS0_)

	.visible .func _Z4SwapIfEvRT_S1_ (.param .u64 __cudaparmf1__Z4SwapIfEvRT_S1_, .param .u64 __cudaparmf2__Z4SwapIfEvRT_S1_)

	.visible .func _Z26ConvertPixelTwo_422_To_4446float4S_14IR_PixelFormatPS_S1_S0_ (.param .align 16 .b8 __cudaparmf1__Z26ConvertPixelTwo_422_To_4446float4S_14IR_PixelFormatPS_S1_S0_[16], .param .align 16 .b8 __cudaparmf2__Z26ConvertPixelTwo_422_To_4446float4S_14IR_PixelFormatPS_S1_S0_[16], .param .s32 __cudaparmf3__Z26ConvertPixelTwo_422_To_4446float4S_14IR_PixelFormatPS_S1_S0_, .param .u64 __cudaparmf4__Z26ConvertPixelTwo_422_To_4446float4S_14IR_PixelFormatPS_S1_S0_, .param .u64 __cudaparmf5__Z26ConvertPixelTwo_422_To_4446float4S_14IR_PixelFormatPS_S1_S0_, .param .s32 __cudaparmf6__Z26ConvertPixelTwo_422_To_4446float4S_14IR_PixelFormatPS_S1_S0_)

	.visible .func _Z33ConvertPixelTwo_422_8u_To_444_32f6uchar4S_14IR_PixelFormatP6float4S2_S0_ (.param .align 4 .b8 __cudaparmf1__Z33ConvertPixelTwo_422_8u_To_444_32f6uchar4S_14IR_PixelFormatP6float4S2_S0_[4], .param .align 4 .b8 __cudaparmf2__Z33ConvertPixelTwo_422_8u_To_444_32f6uchar4S_14IR_PixelFormatP6float4S2_S0_[4], .param .s32 __cudaparmf3__Z33ConvertPixelTwo_422_8u_To_444_32f6uchar4S_14IR_PixelFormatP6float4S2_S0_, .param .u64 __cudaparmf4__Z33ConvertPixelTwo_422_8u_To_444_32f6uchar4S_14IR_PixelFormatP6float4S2_S0_, .param .u64 __cudaparmf5__Z33ConvertPixelTwo_422_8u_To_444_32f6uchar4S_14IR_PixelFormatP6float4S2_S0_, .param .s32 __cudaparmf6__Z33ConvertPixelTwo_422_8u_To_444_32f6uchar4S_14IR_PixelFormatP6float4S2_S0_)

	.visible .func (.param .align 16 .b8 __cudaretf__Z30ConvertPixel_422_8u_To_444_32f6uchar414IR_PixelFormatS0_[16]) _Z30ConvertPixel_422_8u_To_444_32f6uchar414IR_PixelFormatS0_ (.param .align 4 .b8 __cudaparmf1__Z30ConvertPixel_422_8u_To_444_32f6uchar414IR_PixelFormatS0_[4], .param .s32 __cudaparmf2__Z30ConvertPixel_422_8u_To_444_32f6uchar414IR_PixelFormatS0_, .param .s32 __cudaparmf3__Z30ConvertPixel_422_8u_To_444_32f6uchar414IR_PixelFormatS0_)

	.visible .func (.param .align 4 .b8 __cudaretf__Z6Read2DI6uchar4ET_PKS1_iii[4]) _Z6Read2DI6uchar4ET_PKS1_iii (.param .u64 __cudaparmf1__Z6Read2DI6uchar4ET_PKS1_iii, .param .s32 __cudaparmf2__Z6Read2DI6uchar4ET_PKS1_iii, .param .s32 __cudaparmf3__Z6Read2DI6uchar4ET_PKS1_iii, .param .s32 __cudaparmf4__Z6Read2DI6uchar4ET_PKS1_iii)

	.visible .func (.param .align 16 .b8 __cudaretf__Z6Read2DI5uint4ET_PKS1_iii[16]) _Z6Read2DI5uint4ET_PKS1_iii (.param .u64 __cudaparmf1__Z6Read2DI5uint4ET_PKS1_iii, .param .s32 __cudaparmf2__Z6Read2DI5uint4ET_PKS1_iii, .param .s32 __cudaparmf3__Z6Read2DI5uint4ET_PKS1_iii, .param .s32 __cudaparmf4__Z6Read2DI5uint4ET_PKS1_iii)

	//-----------------------------------------------------------
	// Compiling C:/Users/dvaeng/AppData/Local/Temp/tmpxft_00003ca4_00000000-11_PixelFormatConvert_422.cpp3.i (C:/Users/dvaeng/AppData/Local/Temp/ccBI#.a15936)
	//-----------------------------------------------------------

	//-----------------------------------------------------------
	// Options:
	//-----------------------------------------------------------
	//  Target:ptx, ISA:sm_20, Endian:little, Pointer Size:64
	//  -O3	(Optimization level)
	//  -g0	(Debug level)
	//  -m2	(Report advisories)
	//-----------------------------------------------------------

	.file	1	"C:/Users/dvaeng/AppData/Local/Temp/tmpxft_00003ca4_00000000-10_PixelFormatConvert_422.cudafe2.gpu"
	.file	2	"c:\Mulder64\shared\adobe\MediaCore\ImageRenderer\API\Inc\ImageRenderer.h"
	.file	3	"c:\Mulder64\shared\adobe\MediaCore\GPUFoundation\API\Inc\GPUFoundation/PixelFormat.h"
	.file	4	"c:\Mulder64\shared\adobe\MediaCore\GPUFoundation\API\Inc\GPUFoundation/KernelSupport/PixelRGB.h"
	.file	5	"C:\Program Files (x86)\Microsoft Visual Studio 9.0\VC\include\crtdefs.h"
	.file	6	"c:\Mulder64\shared\adobe\MediaCore\External\3rdParty\NVIDIA\CUDA\win\include\crt/device_runtime.h"
	.file	7	"c:\Mulder64\shared\adobe\MediaCore\External\3rdParty\NVIDIA\CUDA\win\include\host_defines.h"
	.file	8	"c:\Mulder64\shared\adobe\MediaCore\External\3rdParty\NVIDIA\CUDA\win\include\builtin_types.h"
	.file	9	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\device_types.h"
	.file	10	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\driver_types.h"
	.file	11	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\surface_types.h"
	.file	12	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\texture_types.h"
	.file	13	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\vector_types.h"
	.file	14	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\builtin_types.h"
	.file	15	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\host_defines.h"
	.file	16	"c:\Mulder64\shared\adobe\MediaCore\External\3rdParty\NVIDIA\CUDA\win\include\device_launch_parameters.h"
	.file	17	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\crt\storage_class.h"
	.file	18	"C:\Program Files (x86)\Microsoft Visual Studio 9.0\VC\include\time.h"
	.file	19	"c:\Mulder64\shared\adobe\MediaCore\GPUFoundation\API\Inc\GPUFoundation/KernelSupport/Utils.h"
	.file	20	"c:\Mulder64\shared\adobe\MediaCore\GPUFoundation\Inc\ImageProcessing/PixelFormatConvert_Common.h"
	.file	21	"c:\Mulder64\shared\adobe\MediaCore\GPUFoundation\API\Inc\GPUFoundation/KernelSupport/VectorUtils.h"
	.file	22	"c:/Mulder64/shared/adobe/MediaCore/GPUFoundation/Src/ImageProcessing/PixelFormatConvert_422.cu"
	.file	23	"c:\Mulder64\shared\adobe\MediaCore\GPUFoundation\API\Inc\GPUFoundation/Numeric.h"
	.file	24	"c:\Mulder64\shared\adobe\MediaCore\External\3rdParty\NVIDIA\CUDA\win\include\common_functions.h"
	.file	25	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\math_functions.h"
	.file	26	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\math_constants.h"
	.file	27	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\device_functions.h"
	.file	28	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\sm_11_atomic_functions.h"
	.file	29	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\sm_12_atomic_functions.h"
	.file	30	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\sm_13_double_functions.h"
	.file	31	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\sm_20_atomic_functions.h"
	.file	32	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\sm_20_intrinsics.h"
	.file	33	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\surface_functions.h"
	.file	34	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\texture_fetch_functions.h"
	.file	35	"c:\mulder64\shared\adobe\mediacore\external\3rdparty\nvidia\cuda\win\include\math_functions_dbl_ptx3.h"
	.file	36	"c:\Mulder64\shared\adobe\MediaCore\GPUFoundation\API\Inc\GPUFoundation/KernelSupport/ColorSpaceConvert.h"


	.visible .func (.param .s32 __cudaretf__Z15IntegerMultiplyii) _Z15IntegerMultiplyii (.param .s32 __cudaparmf1__Z15IntegerMultiplyii, .param .s32 __cudaparmf2__Z15IntegerMultiplyii)
	{
	.reg .u32 %r<7>;
	.loc	19	60	0
$LDWbegin__Z15IntegerMultiplyii:
	ld.param.u32 	%r1, [__cudaparmf1__Z15IntegerMultiplyii];
	mov.s32 	%r2, %r1;
	ld.param.u32 	%r3, [__cudaparmf2__Z15IntegerMultiplyii];
	mov.s32 	%r4, %r3;
	.loc	19	64	0
	mul.lo.s32 	%r5, %r2, %r4;
	st.param.s32 	[__cudaretf__Z15IntegerMultiplyii], %r5;
	ret;
$LDWend__Z15IntegerMultiplyii:
	} // _Z15IntegerMultiplyii

	.visible .func (.param .s32 __cudaretf__Z17Standard2DKernelXv) _Z17Standard2DKernelXv ()
	{
	.reg .u32 %r<7>;
	.loc	19	73	0
$LDWbegin__Z17Standard2DKernelXv:
	.loc	19	74	0
	mov.u32 	%r1, %tid.x;
	cvt.s32.u32 	%r2, %ctaid.x;
	cvt.s32.u32 	%r3, %ntid.x;
	mul.lo.s32 	%r4, %r2, %r3;
	add.u32 	%r5, %r1, %r4;
	st.param.s32 	[__cudaretf__Z17Standard2DKernelXv], %r5;
	ret;
$LDWend__Z17Standard2DKernelXv:
	} // _Z17Standard2DKernelXv

	.visible .func (.param .s32 __cudaretf__Z17Standard2DKernelYv) _Z17Standard2DKernelYv ()
	{
	.reg .u32 %r<7>;
	.loc	19	77	0
$LDWbegin__Z17Standard2DKernelYv:
	.loc	19	78	0
	mov.u32 	%r1, %tid.y;
	cvt.s32.u32 	%r2, %ctaid.y;
	cvt.s32.u32 	%r3, %ntid.y;
	mul.lo.s32 	%r4, %r2, %r3;
	add.u32 	%r5, %r1, %r4;
	st.param.s32 	[__cudaretf__Z17Standard2DKernelYv], %r5;
	ret;
$LDWend__Z17Standard2DKernelYv:
	} // _Z17Standard2DKernelYv

	.visible .func (.param .align 16 .b8 __cudaretf__Z13Half4ToFloat47ushort4[16]) _Z13Half4ToFloat47ushort4 (.param .align 8 .b8 __cudaparmf1__Z13Half4ToFloat47ushort4[8])
	{
	.reg .u32 %r<14>;
	.reg .f32 %f<9>;
	.loc	19	86	0
$LDWbegin__Z13Half4ToFloat47ushort4:
	ld.param.u16 	%r1, [__cudaparmf1__Z13Half4ToFloat47ushort4+0];
	mov.s32 	%r2, %r1;
	ld.param.u16 	%r3, [__cudaparmf1__Z13Half4ToFloat47ushort4+2];
	mov.s32 	%r4, %r3;
	ld.param.u16 	%r5, [__cudaparmf1__Z13Half4ToFloat47ushort4+4];
	mov.s32 	%r6, %r5;
	ld.param.u16 	%r7, [__cudaparmf1__Z13Half4ToFloat47ushort4+6];
	mov.s32 	%r8, %r7;
	.loc	19	87	0
	cvt.u16.u32 	%r9, %r4;
	{ .reg .b32 %b1;
	mov.b32		%b1, %r9;
	cvt.ftz.f32.f16	%f1, %b1; }
	cvt.u16.u32 	%r10, %r6;
	{ .reg .b32 %b1;
	mov.b32		%b1, %r10;
	cvt.ftz.f32.f16	%f2, %b1; }
	cvt.u16.u32 	%r11, %r8;
	{ .reg .b32 %b1;
	mov.b32		%b1, %r11;
	cvt.ftz.f32.f16	%f3, %b1; }
	cvt.u16.u32 	%r12, %r2;
	{ .reg .b32 %b1;
	mov.b32		%b1, %r12;
	cvt.ftz.f32.f16	%f4, %b1; }
	st.param.f32 	[__cudaretf__Z13Half4ToFloat47ushort4+0], %f4;
	mov.f32 	%f5, %f1;
	st.param.f32 	[__cudaretf__Z13Half4ToFloat47ushort4+4], %f5;
	mov.f32 	%f6, %f2;
	st.param.f32 	[__cudaretf__Z13Half4ToFloat47ushort4+8], %f6;
	mov.f32 	%f7, %f3;
	st.param.f32 	[__cudaretf__Z13Half4ToFloat47ushort4+12], %f7;
	ret;
$LDWend__Z13Half4ToFloat47ushort4:
	} // _Z13Half4ToFloat47ushort4

	.visible .func (.param .align 8 .b8 __cudaretf__Z13Float4ToHalf46float4[8]) _Z13Float4ToHalf46float4 (.param .align 16 .b8 __cudaparmf1__Z13Float4ToHalf46float4[16])
	{
	.reg .u32 %r<13>;
	.reg .f32 %f<10>;
	.loc	19	95	0
$LDWbegin__Z13Float4ToHalf46float4:
	ld.param.f32 	%f1, [__cudaparmf1__Z13Float4ToHalf46float4+0];
	mov.f32 	%f2, %f1;
	ld.param.f32 	%f3, [__cudaparmf1__Z13Float4ToHalf46float4+4];
	mov.f32 	%f4, %f3;
	ld.param.f32 	%f5, [__cudaparmf1__Z13Float4ToHalf46float4+8];
	mov.f32 	%f6, %f5;
	ld.param.f32 	%f7, [__cudaparmf1__Z13Float4ToHalf46float4+12];
	mov.f32 	%f8, %f7;
	.loc	19	96	0
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f4;
	mov.b32		%r1, %b1; }
	cvt.u16.u32 	%r2, %r1;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f6;
	mov.b32		%r3, %b1; }
	cvt.u16.u32 	%r4, %r3;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f8;
	mov.b32		%r5, %b1; }
	cvt.u16.u32 	%r6, %r5;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f2;
	mov.b32		%r7, %b1; }
	cvt.u16.u32 	%r8, %r7;
	st.param.u16 	[__cudaretf__Z13Float4ToHalf46float4+0], %r8;
	mov.s32 	%r9, %r2;
	st.param.u16 	[__cudaretf__Z13Float4ToHalf46float4+2], %r9;
	mov.s32 	%r10, %r4;
	st.param.u16 	[__cudaretf__Z13Float4ToHalf46float4+4], %r10;
	mov.s32 	%r11, %r6;
	st.param.u16 	[__cudaretf__Z13Float4ToHalf46float4+6], %r11;
	ret;
$LDWend__Z13Float4ToHalf46float4:
	} // _Z13Float4ToHalf46float4

	.visible .func (.param .u32 __cudaretf__Z4Mix3RjS_S_) _Z4Mix3RjS_S_ (.param .u64 __cudaparmf1__Z4Mix3RjS_S_, .param .u64 __cudaparmf2__Z4Mix3RjS_S_, .param .u64 __cudaparmf3__Z4Mix3RjS_S_)
	{
	.reg .u32 %r<75>;
	.reg .u64 %rd<8>;
	.loc	19	138	0
$LDWbegin__Z4Mix3RjS_S_:
	ld.param.u64 	%rd1, [__cudaparmf1__Z4Mix3RjS_S_];
	mov.s64 	%rd2, %rd1;
	ld.param.u64 	%rd3, [__cudaparmf2__Z4Mix3RjS_S_];
	mov.s64 	%rd4, %rd3;
	ld.param.u64 	%rd5, [__cudaparmf3__Z4Mix3RjS_S_];
	mov.s64 	%rd6, %rd5;
	.loc	19	139	0
	ld.u32 	%r1, [%rd2+0];
	ld.u32 	%r2, [%rd4+0];
	sub.u32 	%r3, %r1, %r2;
	st.u32 	[%rd2+0], %r3;
	ld.u32 	%r4, [%rd6+0];
	sub.u32 	%r5, %r3, %r4;
	st.u32 	[%rd2+0], %r5;
	ld.u32 	%r6, [%rd6+0];
	shr.u32 	%r7, %r6, 13;
	xor.b32 	%r8, %r5, %r7;
	st.u32 	[%rd2+0], %r8;
	.loc	19	140	0
	ld.u32 	%r9, [%rd4+0];
	ld.u32 	%r10, [%rd6+0];
	sub.u32 	%r11, %r9, %r10;
	st.u32 	[%rd4+0], %r11;
	ld.u32 	%r12, [%rd2+0];
	sub.u32 	%r13, %r11, %r12;
	st.u32 	[%rd4+0], %r13;
	ld.u32 	%r14, [%rd2+0];
	shl.b32 	%r15, %r14, 8;
	xor.b32 	%r16, %r13, %r15;
	st.u32 	[%rd4+0], %r16;
	.loc	19	141	0
	ld.u32 	%r17, [%rd6+0];
	ld.u32 	%r18, [%rd2+0];
	sub.u32 	%r19, %r17, %r18;
	st.u32 	[%rd6+0], %r19;
	ld.u32 	%r20, [%rd4+0];
	sub.u32 	%r21, %r19, %r20;
	st.u32 	[%rd6+0], %r21;
	ld.u32 	%r22, [%rd4+0];
	shr.u32 	%r23, %r22, 13;
	xor.b32 	%r24, %r21, %r23;
	st.u32 	[%rd6+0], %r24;
	.loc	19	142	0
	ld.u32 	%r25, [%rd2+0];
	ld.u32 	%r26, [%rd4+0];
	sub.u32 	%r27, %r25, %r26;
	st.u32 	[%rd2+0], %r27;
	ld.u32 	%r28, [%rd6+0];
	sub.u32 	%r29, %r27, %r28;
	st.u32 	[%rd2+0], %r29;
	ld.u32 	%r30, [%rd6+0];
	shr.u32 	%r31, %r30, 12;
	xor.b32 	%r32, %r29, %r31;
	st.u32 	[%rd2+0], %r32;
	.loc	19	143	0
	ld.u32 	%r33, [%rd4+0];
	ld.u32 	%r34, [%rd6+0];
	sub.u32 	%r35, %r33, %r34;
	st.u32 	[%rd4+0], %r35;
	ld.u32 	%r36, [%rd2+0];
	sub.u32 	%r37, %r35, %r36;
	st.u32 	[%rd4+0], %r37;
	ld.u32 	%r38, [%rd2+0];
	shl.b32 	%r39, %r38, 16;
	xor.b32 	%r40, %r37, %r39;
	st.u32 	[%rd4+0], %r40;
	.loc	19	144	0
	ld.u32 	%r41, [%rd6+0];
	ld.u32 	%r42, [%rd2+0];
	sub.u32 	%r43, %r41, %r42;
	st.u32 	[%rd6+0], %r43;
	ld.u32 	%r44, [%rd4+0];
	sub.u32 	%r45, %r43, %r44;
	st.u32 	[%rd6+0], %r45;
	ld.u32 	%r46, [%rd4+0];
	shr.u32 	%r47, %r46, 5;
	xor.b32 	%r48, %r45, %r47;
	st.u32 	[%rd6+0], %r48;
	.loc	19	145	0
	ld.u32 	%r49, [%rd2+0];
	ld.u32 	%r50, [%rd4+0];
	sub.u32 	%r51, %r49, %r50;
	st.u32 	[%rd2+0], %r51;
	ld.u32 	%r52, [%rd6+0];
	sub.u32 	%r53, %r51, %r52;
	st.u32 	[%rd2+0], %r53;
	ld.u32 	%r54, [%rd6+0];
	shr.u32 	%r55, %r54, 3;
	xor.b32 	%r56, %r53, %r55;
	st.u32 	[%rd2+0], %r56;
	.loc	19	146	0
	ld.u32 	%r57, [%rd4+0];
	ld.u32 	%r58, [%rd6+0];
	sub.u32 	%r59, %r57, %r58;
	st.u32 	[%rd4+0], %r59;
	ld.u32 	%r60, [%rd2+0];
	sub.u32 	%r61, %r59, %r60;
	st.u32 	[%rd4+0], %r61;
	ld.u32 	%r62, [%rd2+0];
	shl.b32 	%r63, %r62, 10;
	xor.b32 	%r64, %r61, %r63;
	st.u32 	[%rd4+0], %r64;
	.loc	19	147	0
	ld.u32 	%r65, [%rd6+0];
	ld.u32 	%r66, [%rd2+0];
	sub.u32 	%r67, %r65, %r66;
	st.u32 	[%rd6+0], %r67;
	ld.u32 	%r68, [%rd4+0];
	sub.u32 	%r69, %r67, %r68;
	st.u32 	[%rd6+0], %r69;
	ld.u32 	%r70, [%rd4+0];
	shr.u32 	%r71, %r70, 15;
	xor.b32 	%r72, %r69, %r71;
	st.u32 	[%rd6+0], %r72;
	.loc	19	148	0
	mov.s32 	%r73, %r72;
	st.param.u32 	[__cudaretf__Z4Mix3RjS_S_], %r73;
	ret;
$LDWend__Z4Mix3RjS_S_:
	} // _Z4Mix3RjS_S_

	.visible .func (.param .s32 __cudaretf__Z4Randj) _Z4Randj (.param .u32 __cudaparmf1__Z4Randj)
	{
	.reg .u32 %r<14>;
	.loc	19	152	0
$LDWbegin__Z4Randj:
	ld.param.u32 	%r1, [__cudaparmf1__Z4Randj];
	mov.s32 	%r2, %r1;
	.loc	19	163	0
	mul.lo.u32 	%r3, %r2, 1103515245;
	add.u32 	%r4, %r3, 12345;
	shr.u32 	%r5, %r4, 16;
	and.b32 	%r6, %r5, 255;
	shl.b32 	%r7, %r6, 7;
	mul.lo.u32 	%r8, %r2, -1029531031;
	sub.u32 	%r9, %r8, 740551042;
	shr.u32 	%r10, %r9, 16;
	and.b32 	%r11, %r10, 255;
	xor.b32 	%r12, %r7, %r11;
	st.param.s32 	[__cudaretf__Z4Randj], %r12;
	ret;
$LDWend__Z4Randj:
	} // _Z4Randj

	.visible .func (.param .s32 __cudaretf__Z6Rand2Djjj) _Z6Rand2Djjj (.param .u32 __cudaparmf1__Z6Rand2Djjj, .param .u32 __cudaparmf2__Z6Rand2Djjj, .param .u32 __cudaparmf3__Z6Rand2Djjj)
	{
	.reg .u32 %r<54>;
	.loc	19	169	0
$LDWbegin__Z6Rand2Djjj:
	ld.param.u32 	%r1, [__cudaparmf1__Z6Rand2Djjj];
	mov.s32 	%r2, %r1;
	ld.param.u32 	%r3, [__cudaparmf2__Z6Rand2Djjj];
	mov.s32 	%r4, %r3;
	ld.param.u32 	%r5, [__cudaparmf3__Z6Rand2Djjj];
	mov.s32 	%r6, %r5;
	.loc	19	139	0
	sub.u32 	%r7, %r2, %r4;
	sub.u32 	%r8, %r7, %r6;
	shr.u32 	%r9, %r6, 13;
	xor.b32 	%r10, %r8, %r9;
	.loc	19	140	0
	sub.u32 	%r11, %r4, %r6;
	sub.u32 	%r12, %r11, %r10;
	shl.b32 	%r13, %r10, 8;
	xor.b32 	%r14, %r12, %r13;
	.loc	19	141	0
	sub.u32 	%r15, %r6, %r10;
	sub.u32 	%r16, %r15, %r14;
	shr.u32 	%r17, %r14, 13;
	xor.b32 	%r18, %r16, %r17;
	.loc	19	142	0
	sub.u32 	%r19, %r10, %r14;
	sub.u32 	%r20, %r19, %r18;
	shr.u32 	%r21, %r18, 12;
	xor.b32 	%r22, %r20, %r21;
	.loc	19	143	0
	sub.u32 	%r23, %r14, %r18;
	sub.u32 	%r24, %r23, %r22;
	shl.b32 	%r25, %r22, 16;
	xor.b32 	%r26, %r24, %r25;
	.loc	19	144	0
	sub.u32 	%r27, %r18, %r22;
	sub.u32 	%r28, %r27, %r26;
	shr.u32 	%r29, %r26, 5;
	xor.b32 	%r30, %r28, %r29;
	.loc	19	145	0
	sub.u32 	%r31, %r22, %r26;
	sub.u32 	%r32, %r31, %r30;
	shr.u32 	%r33, %r30, 3;
	xor.b32 	%r34, %r32, %r33;
	.loc	19	146	0
	sub.u32 	%r35, %r26, %r30;
	sub.u32 	%r36, %r35, %r34;
	shl.b32 	%r37, %r34, 10;
	xor.b32 	%r38, %r36, %r37;
	.loc	19	147	0
	sub.u32 	%r39, %r30, %r34;
	sub.u32 	%r40, %r39, %r38;
	shr.u32 	%r41, %r38, 15;
	xor.b32 	%r42, %r40, %r41;
	.loc	19	170	0
	mul.lo.u32 	%r43, %r42, 1103515245;
	add.u32 	%r44, %r43, 12345;
	shr.u32 	%r45, %r44, 16;
	and.b32 	%r46, %r45, 255;
	shl.b32 	%r47, %r46, 7;
	mul.lo.u32 	%r48, %r42, -1029531031;
	sub.u32 	%r49, %r48, 740551042;
	shr.u32 	%r50, %r49, 16;
	and.b32 	%r51, %r50, 255;
	xor.b32 	%r52, %r47, %r51;
	st.param.s32 	[__cudaretf__Z6Rand2Djjj], %r52;
	ret;
$LDWend__Z6Rand2Djjj:
	} // _Z6Rand2Djjj

	.visible .func (.param .s32 __cudaretf__Z6Rand2Dj) _Z6Rand2Dj (.param .u32 __cudaparmf1__Z6Rand2Dj)
	{
	.reg .u32 %r<60>;
	.loc	19	175	0
$LDWbegin__Z6Rand2Dj:
	ld.param.u32 	%r1, [__cudaparmf1__Z6Rand2Dj];
	mov.s32 	%r2, %r1;
	.loc	19	143	0
	cvt.s32.u32 	%r3, %ctaid.y;
	cvt.s32.u32 	%r4, %ntid.y;
	mul.lo.s32 	%r5, %r3, %r4;
	cvt.s32.u32 	%r6, %ctaid.x;
	cvt.s32.u32 	%r7, %ntid.x;
	mul.lo.s32 	%r8, %r6, %r7;
	mov.u32 	%r9, %tid.y;
	add.u32 	%r10, %r5, %r9;
	mov.u32 	%r11, %tid.x;
	add.u32 	%r12, %r8, %r11;
	shr.u32 	%r13, %r10, 13;
	sub.u32 	%r14, %r2, %r12;
	sub.u32 	%r15, %r12, %r10;
	sub.u32 	%r16, %r14, %r10;
	xor.b32 	%r17, %r13, %r16;
	shl.b32 	%r18, %r17, 8;
	sub.u32 	%r19, %r15, %r17;
	sub.u32 	%r20, %r10, %r17;
	xor.b32 	%r21, %r18, %r19;
	shr.u32 	%r22, %r21, 13;
	sub.u32 	%r23, %r20, %r21;
	sub.u32 	%r24, %r17, %r21;
	xor.b32 	%r25, %r22, %r23;
	shr.u32 	%r26, %r25, 12;
	sub.u32 	%r27, %r24, %r25;
	xor.b32 	%r28, %r26, %r27;
	sub.u32 	%r29, %r21, %r25;
	sub.u32 	%r30, %r29, %r28;
	shl.b32 	%r31, %r28, 16;
	xor.b32 	%r32, %r30, %r31;
	.loc	19	144	0
	sub.u32 	%r33, %r25, %r28;
	sub.u32 	%r34, %r33, %r32;
	shr.u32 	%r35, %r32, 5;
	xor.b32 	%r36, %r34, %r35;
	.loc	19	145	0
	sub.u32 	%r37, %r28, %r32;
	sub.u32 	%r38, %r37, %r36;
	shr.u32 	%r39, %r36, 3;
	xor.b32 	%r40, %r38, %r39;
	.loc	19	146	0
	sub.u32 	%r41, %r32, %r36;
	sub.u32 	%r42, %r41, %r40;
	shl.b32 	%r43, %r40, 10;
	xor.b32 	%r44, %r42, %r43;
	.loc	19	147	0
	sub.u32 	%r45, %r36, %r40;
	sub.u32 	%r46, %r45, %r44;
	shr.u32 	%r47, %r44, 15;
	xor.b32 	%r48, %r46, %r47;
	.loc	19	176	0
	mul.lo.u32 	%r49, %r48, 1103515245;
	add.u32 	%r50, %r49, 12345;
	shr.u32 	%r51, %r50, 16;
	and.b32 	%r52, %r51, 255;
	shl.b32 	%r53, %r52, 7;
	mul.lo.u32 	%r54, %r48, -1029531031;
	sub.u32 	%r55, %r54, 740551042;
	shr.u32 	%r56, %r55, 16;
	and.b32 	%r57, %r56, 255;
	xor.b32 	%r58, %r53, %r57;
	st.param.s32 	[__cudaretf__Z6Rand2Dj], %r58;
	ret;
$LDWend__Z6Rand2Dj:
	} // _Z6Rand2Dj

	.visible .func (.param .s32 __cudaretf__Z19MaxUnsignedBitValuei) _Z19MaxUnsignedBitValuei (.param .s32 __cudaparmf1__Z19MaxUnsignedBitValuei)
	{
	.reg .u32 %r<7>;
	.loc	20	51	0
$LDWbegin__Z19MaxUnsignedBitValuei:
	ld.param.u32 	%r1, [__cudaparmf1__Z19MaxUnsignedBitValuei];
	mov.s32 	%r2, %r1;
	.loc	20	52	0
	sub.s32 	%r3, %r2, 1;
	shl.b32 	%r4, 2, %r3;
	sub.s32 	%r5, %r4, 1;
	st.param.s32 	[__cudaretf__Z19MaxUnsignedBitValuei], %r5;
	ret;
$LDWend__Z19MaxUnsignedBitValuei:
	} // _Z19MaxUnsignedBitValuei

	.visible .func (.param .f32 __cudaretf__Z13MaxDepthValue14IR_PixelFormat) _Z13MaxDepthValue14IR_PixelFormat (.param .s32 __cudaparmf1__Z13MaxDepthValue14IR_PixelFormat)
	{
	.reg .u32 %r<9>;
	.reg .f32 %f<4>;
	.reg .pred %p<6>;
	.loc	20	56	0
$LDWbegin__Z13MaxDepthValue14IR_PixelFormat:
	ld.param.u32 	%r1, [__cudaparmf1__Z13MaxDepthValue14IR_PixelFormat];
	mov.s32 	%r2, %r1;
	.loc	20	57	0
	and.b32 	%r3, %r2, 448;
	mov.u32 	%r4, 0;
	setp.ne.s32 	%p1, %r3, %r4;
	@%p1 bra 	$Lt_10_2306;
	.loc	20	59	0
	mov.f32 	%f1, 0f437f0000;     	// 255
	bra.uni 	$LBB10__Z13MaxDepthValue14IR_PixelFormat;
$Lt_10_2306:
	.loc	20	61	0
	mov.u32 	%r5, 64;
	setp.ne.s32 	%p2, %r3, %r5;
	@%p2 bra 	$Lt_10_2562;
	.loc	20	63	0
	mov.f32 	%f1, 0f447fc000;     	// 1023
	bra.uni 	$LBB10__Z13MaxDepthValue14IR_PixelFormat;
$Lt_10_2562:
	.loc	20	65	0
	mov.u32 	%r6, 128;
	setp.ne.s32 	%p3, %r3, %r6;
	@%p3 bra 	$Lt_10_2818;
	.loc	20	68	0
	mov.f32 	%f1, 0f47000000;     	// 32768
	bra.uni 	$LBB10__Z13MaxDepthValue14IR_PixelFormat;
$Lt_10_2818:
	.loc	20	70	0
	mov.u32 	%r7, 192;
	setp.ne.s32 	%p4, %r3, %r7;
	@%p4 bra 	$Lt_10_3074;
	.loc	20	72	0
	mov.f32 	%f1, 0fbf800000;     	// -1
	bra.uni 	$LBB10__Z13MaxDepthValue14IR_PixelFormat;
$Lt_10_3074:
	.loc	20	76	0
	mov.f32 	%f1, 0f3f800000;     	// 1
$LBB10__Z13MaxDepthValue14IR_PixelFormat:
	mov.f32 	%f2, %f1;
	st.param.f32 	[__cudaretf__Z13MaxDepthValue14IR_PixelFormat], %f2;
	ret;
$LDWend__Z13MaxDepthValue14IR_PixelFormat:
	} // _Z13MaxDepthValue14IR_PixelFormat

	.visible .func (.param .f32 __cudaretf__Z15DepthScaleValue14IR_PixelFormatS_) _Z15DepthScaleValue14IR_PixelFormatS_ (.param .s32 __cudaparmf1__Z15DepthScaleValue14IR_PixelFormatS_, .param .s32 __cudaparmf2__Z15DepthScaleValue14IR_PixelFormatS_)
	{
	.reg .u32 %r<16>;
	.reg .f32 %f<5>;
	.reg .pred %p<10>;
	.loc	20	82	0
$LDWbegin__Z15DepthScaleValue14IR_PixelFormatS_:
	ld.param.u32 	%r1, [__cudaparmf1__Z15DepthScaleValue14IR_PixelFormatS_];
	mov.s32 	%r2, %r1;
	ld.param.u32 	%r3, [__cudaparmf2__Z15DepthScaleValue14IR_PixelFormatS_];
	mov.s32 	%r4, %r3;
	.loc	20	57	0
	and.b32 	%r5, %r4, 448;
	mov.u32 	%r6, 0;
	setp.ne.s32 	%p1, %r5, %r6;
	@%p1 bra 	$Lt_11_5378;
	.loc	20	59	0
	mov.f32 	%f1, 0f437f0000;     	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_188_3;
$Lt_11_5378:
	.loc	20	61	0
	mov.u32 	%r7, 64;
	setp.ne.s32 	%p2, %r5, %r7;
	@%p2 bra 	$Lt_11_5634;
	.loc	20	63	0
	mov.f32 	%f1, 0f447fc000;     	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_188_3;
$Lt_11_5634:
	.loc	20	65	0
	mov.u32 	%r8, 128;
	setp.ne.s32 	%p3, %r5, %r8;
	@%p3 bra 	$Lt_11_5890;
	.loc	20	68	0
	mov.f32 	%f1, 0f47000000;     	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_188_3;
$Lt_11_5890:
	.loc	20	70	0
	mov.u32 	%r9, 192;
	setp.ne.s32 	%p4, %r5, %r9;
	@%p4 bra 	$Lt_11_6146;
	.loc	20	72	0
	mov.f32 	%f1, 0fbf800000;     	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_188_3;
$Lt_11_6146:
	.loc	20	76	0
	mov.f32 	%f1, 0f3f800000;     	// 1
$LDWendi__Z19MaxUnsignedBitValuei_188_3:
	.loc	20	57	0
	and.b32 	%r10, %r2, 448;
	mov.u32 	%r11, 0;
	setp.ne.s32 	%p5, %r10, %r11;
	@%p5 bra 	$Lt_11_6402;
	.loc	20	59	0
	mov.f32 	%f2, 0f437f0000;     	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_188_1;
$Lt_11_6402:
	.loc	20	61	0
	mov.u32 	%r12, 64;
	setp.ne.s32 	%p6, %r10, %r12;
	@%p6 bra 	$Lt_11_6658;
	.loc	20	63	0
	mov.f32 	%f2, 0f447fc000;     	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_188_1;
$Lt_11_6658:
	.loc	20	65	0
	mov.u32 	%r13, 128;
	setp.ne.s32 	%p7, %r10, %r13;
	@%p7 bra 	$Lt_11_6914;
	.loc	20	68	0
	mov.f32 	%f2, 0f47000000;     	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_188_1;
$Lt_11_6914:
	.loc	20	70	0
	mov.u32 	%r14, 192;
	setp.ne.s32 	%p8, %r10, %r14;
	@%p8 bra 	$Lt_11_7170;
	.loc	20	72	0
	mov.f32 	%f2, 0fbf800000;     	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_188_1;
$Lt_11_7170:
	.loc	20	76	0
	mov.f32 	%f2, 0f3f800000;     	// 1
$LDWendi__Z19MaxUnsignedBitValuei_188_1:
	.loc	20	83	0
	div.approx.ftz.f32 	%f3, %f1, %f2;
	st.param.f32 	[__cudaretf__Z15DepthScaleValue14IR_PixelFormatS_], %f3;
	ret;
$LDWend__Z15DepthScaleValue14IR_PixelFormatS_:
	} // _Z15DepthScaleValue14IR_PixelFormatS_

	.visible .func (.param .s32 __cudaretf__Z7IsYCbCr14IR_PixelFormat) _Z7IsYCbCr14IR_PixelFormat (.param .s32 __cudaparmf1__Z7IsYCbCr14IR_PixelFormat)
	{
	.reg .u32 %r<10>;
	.loc	20	92	0
$LDWbegin__Z7IsYCbCr14IR_PixelFormat:
	ld.param.u32 	%r1, [__cudaparmf1__Z7IsYCbCr14IR_PixelFormat];
	mov.s32 	%r2, %r1;
	.loc	20	93	0
	and.b32 	%r3, %r2, 448;
	mov.s32 	%r4, 256;
	set.ne.u32.s32 	%r5, %r3, %r4;
	neg.s32 	%r6, %r5;
	and.b32 	%r7, %r2, 1;
	and.b32 	%r8, %r6, %r7;
	st.param.s32 	[__cudaretf__Z7IsYCbCr14IR_PixelFormat], %r8;
	ret;
$LDWend__Z7IsYCbCr14IR_PixelFormat:
	} // _Z7IsYCbCr14IR_PixelFormat
	.const .align 4 .b8 kYCbCrOffset[12] = {0,0,128,65,0,0,0,67,0,0,0,67};
	.const .align 4 .b8 kYCbCrFullRangeOffset[12] = {0,0,0,0,0,0,0,67,0,0,0,67};

	.visible .func (.param .f32 __cudaretf__Z11YCbCrOffseti14IR_PixelFormat) _Z11YCbCrOffseti14IR_PixelFormat (.param .s32 __cudaparmf1__Z11YCbCrOffseti14IR_PixelFormat, .param .s32 __cudaparmf2__Z11YCbCrOffseti14IR_PixelFormat)
	{
	.reg .u32 %r<8>;
	.reg .u64 %rd<8>;
	.reg .f32 %f<4>;
	.reg .pred %p<3>;
	.loc	20	99	0
$LDWbegin__Z11YCbCrOffseti14IR_PixelFormat:
	ld.param.u32 	%r1, [__cudaparmf1__Z11YCbCrOffseti14IR_PixelFormat];
	mov.s32 	%r2, %r1;
	ld.param.u32 	%r3, [__cudaparmf2__Z11YCbCrOffseti14IR_PixelFormat];
	mov.s32 	%r4, %r3;
	cvt.s64.s32 	%rd1, %r2;
	mul.wide.s32 	%rd2, %r2, 4;
	and.b32 	%r5, %r4, 2048;
	mov.u32 	%r6, 0;
	setp.eq.s32 	%p1, %r5, %r6;
	@%p1 bra 	$Lt_13_1282;
	.loc	20	100	0
	mov.u64 	%rd3, kYCbCrOffset;
	add.u64 	%rd4, %rd2, %rd3;
	ld.const.f32 	%f1, [%rd4+0];
	bra.uni 	$Lt_13_1026;
$Lt_13_1282:
	mov.u64 	%rd5, kYCbCrFullRangeOffset;
	add.u64 	%rd6, %rd2, %rd5;
	ld.const.f32 	%f1, [%rd6+0];
$Lt_13_1026:
	mov.f32 	%f2, %f1;
	st.param.f32 	[__cudaretf__Z11YCbCrOffseti14IR_PixelFormat], %f2;
	ret;
$LDWend__Z11YCbCrOffseti14IR_PixelFormat:
	} // _Z11YCbCrOffseti14IR_PixelFormat

	.visible .func (.param .align 16 .b8 __cudaretf__Z15AddAYCbCrOffset6float414IR_PixelFormat[16]) _Z15AddAYCbCrOffset6float414IR_PixelFormat (.param .align 16 .b8 __cudaparmf1__Z15AddAYCbCrOffset6float414IR_PixelFormat[16], .param .s32 __cudaparmf2__Z15AddAYCbCrOffset6float414IR_PixelFormat)
	{
	.reg .u32 %r<11>;
	.reg .f32 %f<23>;
	.reg .pred %p<7>;
	.loc	20	105	0
$LDWbegin__Z15AddAYCbCrOffset6float414IR_PixelFormat:
	ld.param.f32 	%f1, [__cudaparmf1__Z15AddAYCbCrOffset6float414IR_PixelFormat+0];
	mov.f32 	%f2, %f1;
	ld.param.f32 	%f3, [__cudaparmf1__Z15AddAYCbCrOffset6float414IR_PixelFormat+4];
	mov.f32 	%f4, %f3;
	ld.param.f32 	%f5, [__cudaparmf1__Z15AddAYCbCrOffset6float414IR_PixelFormat+8];
	mov.f32 	%f6, %f5;
	ld.param.f32 	%f7, [__cudaparmf1__Z15AddAYCbCrOffset6float414IR_PixelFormat+12];
	mov.f32 	%f8, %f7;
	ld.param.u32 	%r1, [__cudaparmf2__Z15AddAYCbCrOffset6float414IR_PixelFormat];
	mov.s32 	%r2, %r1;
	.loc	20	57	0
	and.b32 	%r3, %r2, 448;
	mov.u32 	%r4, 0;
	setp.ne.s32 	%p1, %r3, %r4;
	@%p1 bra 	$Lt_14_6914;
	.loc	20	59	0
	mov.f32 	%f9, 0f437f0000;     	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_191_3;
$Lt_14_6914:
	.loc	20	61	0
	mov.u32 	%r5, 64;
	setp.ne.s32 	%p2, %r3, %r5;
	@%p2 bra 	$Lt_14_7170;
	.loc	20	63	0
	mov.f32 	%f9, 0f447fc000;     	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_191_3;
$Lt_14_7170:
	.loc	20	65	0
	mov.u32 	%r6, 128;
	setp.ne.s32 	%p3, %r3, %r6;
	@%p3 bra 	$Lt_14_7426;
	.loc	20	68	0
	mov.f32 	%f9, 0f47000000;     	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_191_3;
$Lt_14_7426:
	.loc	20	70	0
	mov.u32 	%r7, 192;
	setp.ne.s32 	%p4, %r3, %r7;
	@%p4 bra 	$Lt_14_7682;
	.loc	20	72	0
	mov.f32 	%f9, 0fbf800000;     	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_191_3;
$Lt_14_7682:
	.loc	20	76	0
	mov.f32 	%f9, 0f3f800000;     	// 1
$LDWendi__Z19MaxUnsignedBitValuei_191_3:
	.loc	20	107	0
	and.b32 	%r8, %r2, 2048;
	mov.s32 	%r9, 0;
	setp.ne.s32 	%p5, %r8, %r9;
	@!%p5 bra 	$Lt_14_11266;
	.loc	20	100	0
	ld.const.f32 	%f10, [kYCbCrOffset+0];
	bra.uni 	$Lt_14_11010;
$Lt_14_11266:
	ld.const.f32 	%f10, [kYCbCrFullRangeOffset+0];
$Lt_14_11010:
	.loc	20	107	0
	mov.f32 	%f11, 0f437f0000;    	// 255
	div.approx.ftz.f32 	%f12, %f9, %f11;
	fma.rn.ftz.f32 	%f13, %f12, %f10, %f4;
	@!%p5 bra 	$Lt_14_11778;
	.loc	20	100	0
	ld.const.f32 	%f14, [kYCbCrOffset+4];
	bra.uni 	$Lt_14_11522;
$Lt_14_11778:
	ld.const.f32 	%f14, [kYCbCrFullRangeOffset+4];
$Lt_14_11522:
	.loc	20	107	0
	fma.rn.ftz.f32 	%f15, %f12, %f14, %f6;
	@!%p5 bra 	$Lt_14_12290;
	.loc	20	100	0
	ld.const.f32 	%f16, [kYCbCrOffset+8];
	bra.uni 	$Lt_14_12034;
$Lt_14_12290:
	ld.const.f32 	%f16, [kYCbCrFullRangeOffset+8];
$Lt_14_12034:
	.loc	20	107	0
	fma.rn.ftz.f32 	%f17, %f12, %f16, %f8;
	mov.f32 	%f18, %f2;
	st.param.f32 	[__cudaretf__Z15AddAYCbCrOffset6float414IR_PixelFormat+0], %f18;
	mov.f32 	%f19, %f13;
	st.param.f32 	[__cudaretf__Z15AddAYCbCrOffset6float414IR_PixelFormat+4], %f19;
	mov.f32 	%f20, %f15;
	st.param.f32 	[__cudaretf__Z15AddAYCbCrOffset6float414IR_PixelFormat+8], %f20;
	mov.f32 	%f21, %f17;
	st.param.f32 	[__cudaretf__Z15AddAYCbCrOffset6float414IR_PixelFormat+12], %f21;
	ret;
$LDWend__Z15AddAYCbCrOffset6float414IR_PixelFormat:
	} // _Z15AddAYCbCrOffset6float414IR_PixelFormat

	.visible .func (.param .align 16 .b8 __cudaretf__Z20SubtractAYCbCrOffset6float414IR_PixelFormat[16]) _Z20SubtractAYCbCrOffset6float414IR_PixelFormat (.param .align 16 .b8 __cudaparmf1__Z20SubtractAYCbCrOffset6float414IR_PixelFormat[16], .param .s32 __cudaparmf2__Z20SubtractAYCbCrOffset6float414IR_PixelFormat)
	{
	.reg .u32 %r<11>;
	.reg .f32 %f<26>;
	.reg .pred %p<7>;
	.loc	20	116	0
$LDWbegin__Z20SubtractAYCbCrOffset6float414IR_PixelFormat:
	ld.param.f32 	%f1, [__cudaparmf1__Z20SubtractAYCbCrOffset6float414IR_PixelFormat+0];
	mov.f32 	%f2, %f1;
	ld.param.f32 	%f3, [__cudaparmf1__Z20SubtractAYCbCrOffset6float414IR_PixelFormat+4];
	mov.f32 	%f4, %f3;
	ld.param.f32 	%f5, [__cudaparmf1__Z20SubtractAYCbCrOffset6float414IR_PixelFormat+8];
	mov.f32 	%f6, %f5;
	ld.param.f32 	%f7, [__cudaparmf1__Z20SubtractAYCbCrOffset6float414IR_PixelFormat+12];
	mov.f32 	%f8, %f7;
	ld.param.u32 	%r1, [__cudaparmf2__Z20SubtractAYCbCrOffset6float414IR_PixelFormat];
	mov.s32 	%r2, %r1;
	.loc	20	57	0
	and.b32 	%r3, %r2, 448;
	mov.u32 	%r4, 0;
	setp.ne.s32 	%p1, %r3, %r4;
	@%p1 bra 	$Lt_15_6914;
	.loc	20	59	0
	mov.f32 	%f9, 0f437f0000;     	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_192_3;
$Lt_15_6914:
	.loc	20	61	0
	mov.u32 	%r5, 64;
	setp.ne.s32 	%p2, %r3, %r5;
	@%p2 bra 	$Lt_15_7170;
	.loc	20	63	0
	mov.f32 	%f9, 0f447fc000;     	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_192_3;
$Lt_15_7170:
	.loc	20	65	0
	mov.u32 	%r6, 128;
	setp.ne.s32 	%p3, %r3, %r6;
	@%p3 bra 	$Lt_15_7426;
	.loc	20	68	0
	mov.f32 	%f9, 0f47000000;     	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_192_3;
$Lt_15_7426:
	.loc	20	70	0
	mov.u32 	%r7, 192;
	setp.ne.s32 	%p4, %r3, %r7;
	@%p4 bra 	$Lt_15_7682;
	.loc	20	72	0
	mov.f32 	%f9, 0fbf800000;     	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_192_3;
$Lt_15_7682:
	.loc	20	76	0
	mov.f32 	%f9, 0f3f800000;     	// 1
$LDWendi__Z19MaxUnsignedBitValuei_192_3:
	.loc	20	118	0
	and.b32 	%r8, %r2, 2048;
	mov.s32 	%r9, 0;
	setp.ne.s32 	%p5, %r8, %r9;
	@!%p5 bra 	$Lt_15_11266;
	.loc	20	100	0
	ld.const.f32 	%f10, [kYCbCrOffset+0];
	bra.uni 	$Lt_15_11010;
$Lt_15_11266:
	ld.const.f32 	%f10, [kYCbCrFullRangeOffset+0];
$Lt_15_11010:
	.loc	20	118	0
	mov.f32 	%f11, 0f437f0000;    	// 255
	div.approx.ftz.f32 	%f12, %f9, %f11;
	mul.ftz.f32 	%f13, %f12, %f10;
	sub.ftz.f32 	%f14, %f4, %f13;
	@!%p5 bra 	$Lt_15_11778;
	.loc	20	100	0
	ld.const.f32 	%f15, [kYCbCrOffset+4];
	bra.uni 	$Lt_15_11522;
$Lt_15_11778:
	ld.const.f32 	%f15, [kYCbCrFullRangeOffset+4];
$Lt_15_11522:
	.loc	20	118	0
	mul.ftz.f32 	%f16, %f12, %f15;
	sub.ftz.f32 	%f17, %f6, %f16;
	@!%p5 bra 	$Lt_15_12290;
	.loc	20	100	0
	ld.const.f32 	%f18, [kYCbCrOffset+8];
	bra.uni 	$Lt_15_12034;
$Lt_15_12290:
	ld.const.f32 	%f18, [kYCbCrFullRangeOffset+8];
$Lt_15_12034:
	.loc	20	118	0
	mul.ftz.f32 	%f19, %f12, %f18;
	sub.ftz.f32 	%f20, %f8, %f19;
	mov.f32 	%f21, %f2;
	st.param.f32 	[__cudaretf__Z20SubtractAYCbCrOffset6float414IR_PixelFormat+0], %f21;
	mov.f32 	%f22, %f14;
	st.param.f32 	[__cudaretf__Z20SubtractAYCbCrOffset6float414IR_PixelFormat+4], %f22;
	mov.f32 	%f23, %f17;
	st.param.f32 	[__cudaretf__Z20SubtractAYCbCrOffset6float414IR_PixelFormat+8], %f23;
	mov.f32 	%f24, %f20;
	st.param.f32 	[__cudaretf__Z20SubtractAYCbCrOffset6float414IR_PixelFormat+12], %f24;
	ret;
$LDWend__Z20SubtractAYCbCrOffset6float414IR_PixelFormat:
	} // _Z20SubtractAYCbCrOffset6float414IR_PixelFormat
	.global .align 4 .b8 kRGB32f_To_601YPbPr[36] = {135,22,153,62,162,69,22,63,213,120,233,61,33,201,44,190,111,155,169,190,0,0,0,63,0,0,0,63,70,94,214,190,232,134,166,189};
	.global .align 4 .b8 k601YPbPr_To_RGB32f[36] = {0,0,128,63,0,0,0,0,188,116,179,63,0,0,128,63,152,50,176,190,158,209,54,191,0,0,128,63,229,208,226,63,0,0,0,0};
	.global .align 4 .b8 kRGB32f_To_601YCbCr[36] = {70,246,130,66,145,141,0,67,94,186,199,65,33,48,23,194,240,103,148,194,0,0,224,66,0,0,224,66,111,146,187,194,70,182,145,193};
	.global .align 4 .b8 k601YCbCr_To_RGB32f[36] = {37,160,149,59,0,0,0,0,182,23,205,59,37,160,149,59,40,15,201,186,156,239,80,187,37,160,149,59,236,155,1,60,0,0,0,0};
	.global .align 4 .b8 kRGB8u_To_601YCbCr[36] = {219,121,131,62,152,14,1,63,18,131,200,61,174,199,23,190,238,252,148,190,197,224,224,62,197,224,224,62,217,78,188,190,174,71,146,189};
	.global .align 4 .b8 k601YCbCr_To_RGB8u[36] = {127,10,149,63,0,0,0,0,160,74,204,63,127,10,149,63,254,148,200,190,184,30,80,191,127,10,149,63,78,26,1,64,0,0,0,0};
	.global .align 4 .b8 kRGB8u_To_601YCbCrFullRange[36] = {135,22,153,62,162,69,22,63,213,120,233,61,166,27,44,190,39,241,168,190,250,254,254,62,250,254,254,62,43,135,213,190,59,223,165,189};
	.global .align 4 .b8 k601YCbCrFullRange_To_RGB8u[36] = {0,0,128,63,0,0,0,0,72,193,178,63,0,0,128,63,143,130,175,190,225,26,54,191,0,0,128,63,20,238,225,63,0,0,0,0};
	.global .align 4 .b8 kRGB32f_To_601YCbCrFullRange[36] = {113,125,152,66,92,175,21,67,92,143,232,65,158,111,43,194,49,72,168,194,0,0,254,66,0,0,254,66,170,177,212,194,88,57,165,193};
	.global .align 4 .b8 k601YCbCrFullRange_To_RGB32f[36] = {129,128,128,59,0,0,0,0,188,116,179,59,129,128,128,59,194,50,176,186,179,209,54,187,129,128,128,59,229,208,226,59,0,0,0,0};
	.global .align 4 .b8 kRGB32f_To_709YPbPr[36] = {208,179,89,62,89,23,55,63,152,221,147,61,186,164,234,189,210,86,197,190,0,0,0,63,0,0,0,63,190,134,232,190,16,202,59,189};
	.global .align 4 .b8 k709YPbPr_To_RGB32f[36] = {0,0,128,63,0,0,0,0,12,147,201,63,0,0,128,63,221,209,63,190,243,173,239,190,0,0,128,63,77,132,237,63,0,0,0,0};
	.global .align 4 .b8 kRGB32f_To_709YCbCr[36] = {106,60,58,66,6,161,28,67,244,253,124,65,223,79,205,193,8,172,172,194,0,0,224,66,0,0,224,66,195,117,203,194,236,81,36,193};
	.global .align 4 .b8 k709YCbCr_To_RGB32f[36] = {37,160,149,59,0,0,0,0,239,94,230,59,37,160,149,59,33,57,91,186,178,245,8,187,37,160,149,59,82,185,7,60,0,0,0,0};
	.global .align 4 .b8 kRGB8u_To_709YCbCr[36] = {207,247,58,62,53,62,29,63,231,251,125,61,147,24,206,61,23,89,173,190,197,224,224,62,197,224,224,62,12,66,204,190,195,245,36,189};
	.global .align 4 .b8 k709YCbCr_To_RGB8u[36] = {127,10,149,63,0,0,0,0,147,120,229,63,127,10,149,63,53,94,90,190,205,108,8,191,127,10,149,63,154,49,7,64,0,0,0,0};
	.global .align 4 .b8 k709YCbCr_To_601YCbCr[36] = {0,0,128,63,23,100,203,61,1,77,68,62,0,0,0,0,18,103,125,63,10,158,226,189,0,0,0,0,61,98,148,189,249,191,123,63};
	.global .align 4 .b8 k601YCbCr_To_709YCbCr[36] = {0,0,128,63,122,165,236,189,179,237,84,190,0,0,0,0,204,98,130,63,216,188,234,61,0,0,0,0,74,179,153,61,234,61,131,63};

	.visible .func (.param .u64 __cudaretf__Z23ColorSpaceConvertMatrix14IR_PixelFormatS_) _Z23ColorSpaceConvertMatrix14IR_PixelFormatS_ (.param .s32 __cudaparmf1__Z23ColorSpaceConvertMatrix14IR_PixelFormatS_, .param .s32 __cudaparmf2__Z23ColorSpaceConvertMatrix14IR_PixelFormatS_)
	{
	.reg .u32 %r<53>;
	.reg .u64 %rd<4>;
	.reg .pred %p<19>;
	.loc	20	132	0
$LDWbegin__Z23ColorSpaceConvertMatrix14IR_PixelFormatS_:
	ld.param.u32 	%r1, [__cudaparmf1__Z23ColorSpaceConvertMatrix14IR_PixelFormatS_];
	mov.s32 	%r2, %r1;
	ld.param.u32 	%r3, [__cudaparmf2__Z23ColorSpaceConvertMatrix14IR_PixelFormatS_];
	mov.s32 	%r4, %r3;
	and.b32 	%r5, %r2, 1;
	mov.u32 	%r6, 0;
	setp.ne.s32 	%p1, %r5, %r6;
	@%p1 bra 	$Lt_16_30210;
	and.b32 	%r7, %r4, 448;
	mov.s32 	%r8, 256;
	setp.eq.s32 	%p2, %r7, %r8;
	and.b32 	%r9, %r2, 448;
	mov.u32 	%r10, 256;
	setp.ne.s32 	%p3, %r9, %r10;
	@%p3 bra 	$Lt_16_30722;
	.loc	20	137	0
	and.b32 	%r11, %r4, 1536;
	mov.s32 	%r12, 512;
	setp.eq.s32 	%p4, %r11, %r12;
	@!%p2 bra 	$Lt_16_18690;
	.loc	20	139	0
	@!%p4 bra 	$Lt_16_18946;
	.loc	20	141	0
	cvta.global.u64 	%rd1, kRGB32f_To_709YPbPr;
	bra.uni 	$LBB53__Z23ColorSpaceConvertMatrix14IR_PixelFormatS_;
$Lt_16_18946:
	.loc	20	145	0
	cvta.global.u64 	%rd1, kRGB32f_To_601YPbPr;
	bra.uni 	$LBB53__Z23ColorSpaceConvertMatrix14IR_PixelFormatS_;
$Lt_16_18690:
	.loc	20	150	0
	@!%p4 bra 	$Lt_16_19202;
	.loc	20	152	0
	cvta.global.u64 	%rd1, kRGB32f_To_709YCbCr;
	bra.uni 	$LBB53__Z23ColorSpaceConvertMatrix14IR_PixelFormatS_;
$Lt_16_19202:
	.loc	20	154	0
	and.b32 	%r13, %r4, 2048;
	mov.u32 	%r14, 0;
	setp.ne.s32 	%p5, %r13, %r14;
	@%p5 bra 	$Lt_16_19458;
	.loc	20	156	0
	cvta.global.u64 	%rd1, kRGB32f_To_601YCbCrFullRange;
	bra.uni 	$LBB53__Z23ColorSpaceConvertMatrix14IR_PixelFormatS_;
$Lt_16_19458:
	.loc	20	160	0
	cvta.global.u64 	%rd1, kRGB32f_To_601YCbCr;
	bra.uni 	$LBB53__Z23ColorSpaceConvertMatrix14IR_PixelFormatS_;
$Lt_16_30722:
	@!%p2 bra 	$Lt_16_31234;
	bra.uni 	$Lt_16_29954;
$Lt_16_31234:
	.loc	20	179	0
	and.b32 	%r15, %r4, 1536;
	mov.u32 	%r16, 512;
	setp.ne.s32 	%p6, %r15, %r16;
	@%p6 bra 	$Lt_16_20226;
	.loc	20	181	0
	cvta.global.u64 	%rd1, kRGB8u_To_709YCbCr;
	bra.uni 	$LBB53__Z23ColorSpaceConvertMatrix14IR_PixelFormatS_;
$Lt_16_20226:
	.loc	20	183	0
	and.b32 	%r17, %r4, 2048;
	mov.u32 	%r18, 0;
	setp.ne.s32 	%p7, %r17, %r18;
	@%p7 bra 	$Lt_16_20482;
	.loc	20	185	0
	cvta.global.u64 	%rd1, kRGB8u_To_601YCbCrFullRange;
	bra.uni 	$LBB53__Z23ColorSpaceConvertMatrix14IR_PixelFormatS_;
$Lt_16_20482:
	.loc	20	189	0
	cvta.global.u64 	%rd1, kRGB8u_To_601YCbCr;
	bra.uni 	$LBB53__Z23ColorSpaceConvertMatrix14IR_PixelFormatS_;
$Lt_16_30210:
	and.b32 	%r19, %r4, 1;
	mov.s32 	%r20, 0;
	setp.eq.s32 	%p8, %r19, %r20;
	and.b32 	%r21, %r2, 1536;
	mov.u32 	%r22, 512;
	setp.ne.s32 	%p9, %r21, %r22;
	@%p9 bra 	$Lt_16_31746;
	and.b32 	%r23, %r2, 448;
	mov.s32 	%r24, 256;
	setp.eq.s32 	%p10, %r23, %r24;
	@!%p8 bra 	$Lt_16_32258;
	and.b32 	%r25, %r4, 448;
	mov.s32 	%r26, 256;
	setp.eq.s32 	%p11, %r25, %r26;
	@!%p10 bra 	$Lt_16_32770;
	@!%p11 bra 	$Lt_16_29954;
	.loc	20	202	0
	cvta.global.u64 	%rd1, k709YPbPr_To_RGB32f;
	bra.uni 	$LBB53__Z23ColorSpaceConvertMatrix14IR_PixelFormatS_;
$Lt_16_32770:
	.loc	20	211	0
	@!%p11 bra 	$Lt_16_21762;
	.loc	20	213	0
	cvta.global.u64 	%rd1, k709YCbCr_To_RGB32f;
	bra.uni 	$LBB53__Z23ColorSpaceConvertMatrix14IR_PixelFormatS_;
$Lt_16_21762:
	.loc	20	217	0
	cvta.global.u64 	%rd1, k709YCbCr_To_RGB8u;
	bra.uni 	$LBB53__Z23ColorSpaceConvertMatrix14IR_PixelFormatS_;
$Lt_16_32258:
	@!%p10 bra 	$Lt_16_33794;
	bra.uni 	$Lt_16_29954;
$Lt_16_33794:
	and.b32 	%r27, %r4, 448;
	mov.s32 	%r28, 256;
	set.eq.u32.s32 	%r29, %r27, %r28;
	neg.s32 	%r30, %r29;
	and.b32 	%r31, %r4, 2048;
	mov.s32 	%r32, 0;
	set.eq.u32.s32 	%r33, %r31, %r32;
	neg.s32 	%r34, %r33;
	or.b32 	%r35, %r30, %r34;
	mov.u32 	%r36, 0;
	setp.eq.s32 	%p12, %r35, %r36;
	@%p12 bra 	$Lt_16_34306;
	bra.uni 	$Lt_16_29954;
$Lt_16_34306:
	.loc	20	250	0
	cvta.global.u64 	%rd1, k709YCbCr_To_601YCbCr;
	bra.uni 	$LBB53__Z23ColorSpaceConvertMatrix14IR_PixelFormatS_;
$Lt_16_31746:
	and.b32 	%r37, %r2, 2048;
	mov.s32 	%r38, 0;
	setp.eq.s32 	%p13, %r37, %r38;
	@!%p13 bra 	$Lt_16_34818;
	@!%p8 bra 	$Lt_16_29954;
	.loc	20	259	0
	and.b32 	%r39, %r4, 448;
	mov.u32 	%r40, 256;
	setp.ne.s32 	%p14, %r39, %r40;
	@%p14 bra 	$Lt_16_23298;
	.loc	20	261	0
	cvta.global.u64 	%rd1, k601YCbCrFullRange_To_RGB32f;
	bra.uni 	$LBB53__Z23ColorSpaceConvertMatrix14IR_PixelFormatS_;
$Lt_16_23298:
	.loc	20	265	0
	cvta.global.u64 	%rd1, k601YCbCrFullRange_To_RGB8u;
	bra.uni 	$LBB53__Z23ColorSpaceConvertMatrix14IR_PixelFormatS_;
$Lt_16_34818:
	and.b32 	%r41, %r2, 448;
	mov.s32 	%r42, 256;
	setp.eq.s32 	%p15, %r41, %r42;
	@!%p8 bra 	$Lt_16_35842;
	and.b32 	%r43, %r4, 448;
	mov.s32 	%r44, 256;
	setp.eq.s32 	%p16, %r43, %r44;
	@!%p15 bra 	$Lt_16_36354;
	@!%p16 bra 	$Lt_16_29954;
	.loc	20	302	0
	cvta.global.u64 	%rd1, k601YPbPr_To_RGB32f;
	bra.uni 	$LBB53__Z23ColorSpaceConvertMatrix14IR_PixelFormatS_;
$Lt_16_36354:
	.loc	20	311	0
	@!%p16 bra 	$Lt_16_25090;
	.loc	20	313	0
	cvta.global.u64 	%rd1, k601YCbCr_To_RGB32f;
	bra.uni 	$LBB53__Z23ColorSpaceConvertMatrix14IR_PixelFormatS_;
$Lt_16_25090:
	.loc	20	317	0
	cvta.global.u64 	%rd1, k601YCbCr_To_RGB8u;
	bra.uni 	$LBB53__Z23ColorSpaceConvertMatrix14IR_PixelFormatS_;
$Lt_16_35842:
	@!%p15 bra 	$Lt_16_37378;
	bra.uni 	$Lt_16_29954;
$Lt_16_37378:
	selp.s32 	%r45, 1, 0, %p13;
	and.b32 	%r46, %r4, 448;
	mov.s32 	%r47, 256;
	set.eq.u32.s32 	%r48, %r46, %r47;
	neg.s32 	%r49, %r48;
	or.b32 	%r50, %r45, %r49;
	mov.u32 	%r51, 0;
	setp.eq.s32 	%p17, %r50, %r51;
	@%p17 bra 	$Lt_16_37890;
	bra.uni 	$Lt_16_29954;
$Lt_16_37890:
	.loc	20	350	0
	cvta.global.u64 	%rd1, k601YCbCr_To_709YCbCr;
	bra.uni 	$LBB53__Z23ColorSpaceConvertMatrix14IR_PixelFormatS_;
$Lt_16_29954:
	.loc	20	355	0
	mov.u64 	%rd1, 0;
$LBB53__Z23ColorSpaceConvertMatrix14IR_PixelFormatS_:
	mov.s64 	%rd2, %rd1;
	st.param.u64 	[__cudaretf__Z23ColorSpaceConvertMatrix14IR_PixelFormatS_], %rd2;
	ret;
$LDWend__Z23ColorSpaceConvertMatrix14IR_PixelFormatS_:
	} // _Z23ColorSpaceConvertMatrix14IR_PixelFormatS_

	.visible .func (.param .s32 __cudaretf__Z13IsGammaLinear14IR_PixelFormat) _Z13IsGammaLinear14IR_PixelFormat (.param .s32 __cudaparmf1__Z13IsGammaLinear14IR_PixelFormat)
	{
	.reg .u32 %r<8>;
	.loc	20	363	0
$LDWbegin__Z13IsGammaLinear14IR_PixelFormat:
	ld.param.u32 	%r1, [__cudaparmf1__Z13IsGammaLinear14IR_PixelFormat];
	mov.s32 	%r2, %r1;
	.loc	20	364	0
	and.b32 	%r3, %r2, 2;
	mov.s32 	%r4, 0;
	set.ne.u32.s32 	%r5, %r3, %r4;
	neg.s32 	%r6, %r5;
	st.param.s32 	[__cudaretf__Z13IsGammaLinear14IR_PixelFormat], %r6;
	ret;
$LDWend__Z13IsGammaLinear14IR_PixelFormat:
	} // _Z13IsGammaLinear14IR_PixelFormat

	.visible .func (.param .f32 __cudaretf__Z18ApplyGammaFunctionff) _Z18ApplyGammaFunctionff (.param .f32 __cudaparmf1__Z18ApplyGammaFunctionff, .param .f32 __cudaparmf2__Z18ApplyGammaFunctionff)
	{
	.reg .f32 %f<15>;
	.reg .pred %p<3>;
	.loc	20	369	0
$LDWbegin__Z18ApplyGammaFunctionff:
	ld.param.f32 	%f1, [__cudaparmf1__Z18ApplyGammaFunctionff];
	mov.f32 	%f2, %f1;
	ld.param.f32 	%f3, [__cudaparmf2__Z18ApplyGammaFunctionff];
	mov.f32 	%f4, %f3;
	mov.f32 	%f5, 0f00000000;     	// 0
	setp.lt.ftz.f32 	%p1, %f2, %f5;
	@!%p1 bra 	$Lt_18_1026;
	.loc	20	372	0
	neg.ftz.f32 	%f6, %f2;
	lg2.approx.ftz.f32 	%f7, %f6;
	mul.ftz.f32 	%f8, %f4, %f7;
	ex2.approx.ftz.f32 	%f9, %f8;
	neg.ftz.f32 	%f10, %f9;
	bra.uni 	$LBB4__Z18ApplyGammaFunctionff;
$Lt_18_1026:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f11, %f2;
	mul.ftz.f32 	%f12, %f4, %f11;
	ex2.approx.ftz.f32 	%f10, %f12;
$LBB4__Z18ApplyGammaFunctionff:
	mov.f32 	%f13, %f10;
	st.param.f32 	[__cudaretf__Z18ApplyGammaFunctionff], %f13;
	ret;
$LDWend__Z18ApplyGammaFunctionff:
	} // _Z18ApplyGammaFunctionff

	.visible .func (.param .align 16 .b8 __cudaretf__Z16OpaqueComponents6float414IR_PixelFormat[16]) _Z16OpaqueComponents6float414IR_PixelFormat (.param .align 16 .b8 __cudaparmf1__Z16OpaqueComponents6float414IR_PixelFormat[16], .param .s32 __cudaparmf2__Z16OpaqueComponents6float414IR_PixelFormat)
	{
	.reg .u32 %r<9>;
	.reg .f32 %f<13>;
	.reg .pred %p<6>;
	.loc	20	401	0
$LDWbegin__Z16OpaqueComponents6float414IR_PixelFormat:
	ld.param.f32 	%f1, [__cudaparmf1__Z16OpaqueComponents6float414IR_PixelFormat+4];
	mov.f32 	%f2, %f1;
	ld.param.f32 	%f3, [__cudaparmf1__Z16OpaqueComponents6float414IR_PixelFormat+8];
	mov.f32 	%f4, %f3;
	ld.param.f32 	%f5, [__cudaparmf1__Z16OpaqueComponents6float414IR_PixelFormat+12];
	mov.f32 	%f6, %f5;
	ld.param.u32 	%r1, [__cudaparmf2__Z16OpaqueComponents6float414IR_PixelFormat];
	mov.s32 	%r2, %r1;
	.loc	20	57	0
	and.b32 	%r3, %r2, 448;
	mov.u32 	%r4, 0;
	setp.ne.s32 	%p1, %r3, %r4;
	@%p1 bra 	$Lt_19_2818;
	.loc	20	59	0
	mov.f32 	%f7, 0f437f0000;     	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_196_1;
$Lt_19_2818:
	.loc	20	61	0
	mov.u32 	%r5, 64;
	setp.ne.s32 	%p2, %r3, %r5;
	@%p2 bra 	$Lt_19_3074;
	.loc	20	63	0
	mov.f32 	%f7, 0f447fc000;     	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_196_1;
$Lt_19_3074:
	.loc	20	65	0
	mov.u32 	%r6, 128;
	setp.ne.s32 	%p3, %r3, %r6;
	@%p3 bra 	$Lt_19_3330;
	.loc	20	68	0
	mov.f32 	%f7, 0f47000000;     	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_196_1;
$Lt_19_3330:
	.loc	20	70	0
	mov.u32 	%r7, 192;
	setp.ne.s32 	%p4, %r3, %r7;
	@%p4 bra 	$Lt_19_3586;
	.loc	20	72	0
	mov.f32 	%f7, 0fbf800000;     	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_196_1;
$Lt_19_3586:
	.loc	20	76	0
	mov.f32 	%f7, 0f3f800000;     	// 1
$LDWendi__Z19MaxUnsignedBitValuei_196_1:
	.loc	20	404	0
	mov.f32 	%f8, %f7;
	st.param.f32 	[__cudaretf__Z16OpaqueComponents6float414IR_PixelFormat+0], %f8;
	mov.f32 	%f9, %f2;
	st.param.f32 	[__cudaretf__Z16OpaqueComponents6float414IR_PixelFormat+4], %f9;
	mov.f32 	%f10, %f4;
	st.param.f32 	[__cudaretf__Z16OpaqueComponents6float414IR_PixelFormat+8], %f10;
	mov.f32 	%f11, %f6;
	st.param.f32 	[__cudaretf__Z16OpaqueComponents6float414IR_PixelFormat+12], %f11;
	ret;
$LDWend__Z16OpaqueComponents6float414IR_PixelFormat:
	} // _Z16OpaqueComponents6float414IR_PixelFormat

	.visible .func (.param .align 16 .b8 __cudaretf__Z21PremultiplyComponents6float414IR_PixelFormat[16]) _Z21PremultiplyComponents6float414IR_PixelFormat (.param .align 16 .b8 __cudaparmf1__Z21PremultiplyComponents6float414IR_PixelFormat[16], .param .s32 __cudaparmf2__Z21PremultiplyComponents6float414IR_PixelFormat)
	{
	.reg .u32 %r<26>;
	.reg .f32 %f<35>;
	.reg .pred %p<16>;
	.loc	20	409	0
$LDWbegin__Z21PremultiplyComponents6float414IR_PixelFormat:
	ld.param.f32 	%f1, [__cudaparmf1__Z21PremultiplyComponents6float414IR_PixelFormat+0];
	mov.f32 	%f2, %f1;
	ld.param.f32 	%f3, [__cudaparmf1__Z21PremultiplyComponents6float414IR_PixelFormat+4];
	mov.f32 	%f4, %f3;
	ld.param.f32 	%f5, [__cudaparmf1__Z21PremultiplyComponents6float414IR_PixelFormat+8];
	mov.f32 	%f6, %f5;
	ld.param.f32 	%f7, [__cudaparmf1__Z21PremultiplyComponents6float414IR_PixelFormat+12];
	mov.f32 	%f8, %f7;
	ld.param.u32 	%r1, [__cudaparmf2__Z21PremultiplyComponents6float414IR_PixelFormat];
	mov.s32 	%r2, %r1;
	.loc	20	410	0
	mov.f32 	%f9, %f4;
	mov.f32 	%f10, %f6;
	mov.f32 	%f11, %f8;
	.loc	20	57	0
	and.b32 	%r3, %r2, 448;
	mov.s32 	%r4, 0;
	setp.eq.s32 	%p1, %r3, %r4;
	@!%p1 bra 	$Lt_20_18690;
	.loc	20	59	0
	mov.f32 	%f12, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_197_9;
$Lt_20_18690:
	.loc	20	61	0
	mov.u32 	%r5, 64;
	setp.ne.s32 	%p2, %r3, %r5;
	@%p2 bra 	$Lt_20_18946;
	.loc	20	63	0
	mov.f32 	%f12, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_197_9;
$Lt_20_18946:
	.loc	20	65	0
	mov.u32 	%r6, 128;
	setp.ne.s32 	%p3, %r3, %r6;
	@%p3 bra 	$Lt_20_19202;
	.loc	20	68	0
	mov.f32 	%f12, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_197_9;
$Lt_20_19202:
	.loc	20	70	0
	mov.u32 	%r7, 192;
	setp.ne.s32 	%p4, %r3, %r7;
	@%p4 bra 	$Lt_20_19458;
	.loc	20	72	0
	mov.f32 	%f12, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_197_9;
$Lt_20_19458:
	.loc	20	76	0
	mov.f32 	%f12, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_197_9:
	.loc	20	413	0
	mov.s32 	%r8, %r2;
	mov.s32 	%r9, 256;
	setp.ne.s32 	%p5, %r3, %r9;
	and.b32 	%r10, %r8, 1;
	selp.s32 	%r11, 1, 0, %p5;
	and.b32 	%r12, %r10, %r11;
	mov.u32 	%r13, 0;
	setp.eq.s32 	%p6, %r12, %r13;
	@%p6 bra 	$Lt_20_29442;
	.loc	20	57	0
	@!%p1 bra 	$Lt_20_19970;
	.loc	20	59	0
	mov.f32 	%f13, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_197_7;
$Lt_20_19970:
	.loc	20	61	0
	mov.u32 	%r14, 64;
	setp.ne.s32 	%p7, %r3, %r14;
	@%p7 bra 	$Lt_20_20226;
	.loc	20	63	0
	mov.f32 	%f13, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_197_7;
$Lt_20_20226:
	.loc	20	65	0
	mov.u32 	%r15, 128;
	setp.ne.s32 	%p8, %r3, %r15;
	@%p8 bra 	$Lt_20_20482;
	.loc	20	68	0
	mov.f32 	%f13, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_197_7;
$Lt_20_20482:
	.loc	20	70	0
	mov.u32 	%r16, 192;
	setp.ne.s32 	%p9, %r3, %r16;
	@%p9 bra 	$Lt_20_20738;
	.loc	20	72	0
	mov.f32 	%f13, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_197_7;
$Lt_20_20738:
	.loc	20	76	0
	mov.f32 	%f13, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_197_7:
	.loc	20	118	0
	and.b32 	%r17, %r2, 2048;
	mov.s32 	%r18, 0;
	setp.ne.s32 	%p10, %r17, %r18;
	@!%p10 bra 	$Lt_20_30210;
	.loc	20	100	0
	ld.const.f32 	%f14, [kYCbCrOffset+0];
	bra.uni 	$Lt_20_29954;
$Lt_20_30210:
	ld.const.f32 	%f14, [kYCbCrFullRangeOffset+0];
$Lt_20_29954:
	.loc	20	118	0
	@!%p10 bra 	$Lt_20_30722;
	.loc	20	100	0
	ld.const.f32 	%f15, [kYCbCrOffset+4];
	bra.uni 	$Lt_20_30466;
$Lt_20_30722:
	ld.const.f32 	%f15, [kYCbCrFullRangeOffset+4];
$Lt_20_30466:
	.loc	20	118	0
	@!%p10 bra 	$Lt_20_31234;
	.loc	20	100	0
	ld.const.f32 	%f16, [kYCbCrOffset+8];
	bra.uni 	$Lt_20_30978;
$Lt_20_31234:
	ld.const.f32 	%f16, [kYCbCrFullRangeOffset+8];
$Lt_20_30978:
	.loc	20	415	0
	mov.f32 	%f17, 0f437f0000;    	// 255
	div.approx.ftz.f32 	%f18, %f13, %f17;
	mul.ftz.f32 	%f19, %f18, %f14;
	sub.ftz.f32 	%f9, %f4, %f19;
	mul.ftz.f32 	%f20, %f18, %f15;
	sub.ftz.f32 	%f10, %f6, %f20;
	mul.ftz.f32 	%f21, %f18, %f16;
	sub.ftz.f32 	%f11, %f8, %f21;
$Lt_20_29442:
	.loc	20	418	0
	rcp.approx.ftz.f32 	%f22, %f12;
	mul.ftz.f32 	%f23, %f22, %f2;
	mul.ftz.f32 	%f9, %f23, %f9;
	.loc	20	419	0
	mul.ftz.f32 	%f10, %f23, %f10;
	.loc	20	420	0
	mul.ftz.f32 	%f11, %f23, %f11;
	.loc	20	422	0
	mov.u32 	%r19, 0;
	setp.eq.s32 	%p11, %r12, %r19;
	@%p11 bra 	$Lt_20_31490;
	.loc	20	57	0
	@!%p1 bra 	$Lt_20_23042;
	.loc	20	59	0
	mov.f32 	%f24, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_197_3;
$Lt_20_23042:
	.loc	20	61	0
	mov.u32 	%r20, 64;
	setp.ne.s32 	%p12, %r3, %r20;
	@%p12 bra 	$Lt_20_23298;
	.loc	20	63	0
	mov.f32 	%f24, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_197_3;
$Lt_20_23298:
	.loc	20	65	0
	mov.u32 	%r21, 128;
	setp.ne.s32 	%p13, %r3, %r21;
	@%p13 bra 	$Lt_20_23554;
	.loc	20	68	0
	mov.f32 	%f24, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_197_3;
$Lt_20_23554:
	.loc	20	70	0
	mov.u32 	%r22, 192;
	setp.ne.s32 	%p14, %r3, %r22;
	@%p14 bra 	$Lt_20_23810;
	.loc	20	72	0
	mov.f32 	%f24, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_197_3;
$Lt_20_23810:
	.loc	20	76	0
	mov.f32 	%f24, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_197_3:
	.loc	20	107	0
	and.b32 	%r23, %r2, 2048;
	mov.s32 	%r24, 0;
	setp.ne.s32 	%p10, %r23, %r24;
	@!%p10 bra 	$Lt_20_32258;
	.loc	20	100	0
	ld.const.f32 	%f25, [kYCbCrOffset+0];
	bra.uni 	$Lt_20_32002;
$Lt_20_32258:
	ld.const.f32 	%f25, [kYCbCrFullRangeOffset+0];
$Lt_20_32002:
	.loc	20	107	0
	@!%p10 bra 	$Lt_20_32770;
	.loc	20	100	0
	ld.const.f32 	%f26, [kYCbCrOffset+4];
	bra.uni 	$Lt_20_32514;
$Lt_20_32770:
	ld.const.f32 	%f26, [kYCbCrFullRangeOffset+4];
$Lt_20_32514:
	.loc	20	107	0
	@!%p10 bra 	$Lt_20_33282;
	.loc	20	100	0
	ld.const.f32 	%f27, [kYCbCrOffset+8];
	bra.uni 	$Lt_20_33026;
$Lt_20_33282:
	ld.const.f32 	%f27, [kYCbCrFullRangeOffset+8];
$Lt_20_33026:
	.loc	20	424	0
	mov.f32 	%f28, 0f437f0000;    	// 255
	div.approx.ftz.f32 	%f29, %f24, %f28;
	fma.rn.ftz.f32 	%f9, %f29, %f25, %f9;
	fma.rn.ftz.f32 	%f10, %f29, %f26, %f10;
	fma.rn.ftz.f32 	%f11, %f29, %f27, %f11;
$Lt_20_31490:
	.loc	20	427	0
	mov.f32 	%f30, %f2;
	st.param.f32 	[__cudaretf__Z21PremultiplyComponents6float414IR_PixelFormat+0], %f30;
	mov.f32 	%f31, %f9;
	st.param.f32 	[__cudaretf__Z21PremultiplyComponents6float414IR_PixelFormat+4], %f31;
	mov.f32 	%f32, %f10;
	st.param.f32 	[__cudaretf__Z21PremultiplyComponents6float414IR_PixelFormat+8], %f32;
	mov.f32 	%f33, %f11;
	st.param.f32 	[__cudaretf__Z21PremultiplyComponents6float414IR_PixelFormat+12], %f33;
	ret;
$LDWend__Z21PremultiplyComponents6float414IR_PixelFormat:
	} // _Z21PremultiplyComponents6float414IR_PixelFormat

	.visible .func (.param .align 16 .b8 __cudaretf__Z23UnpremultiplyComponents6float414IR_PixelFormat[16]) _Z23UnpremultiplyComponents6float414IR_PixelFormat (.param .align 16 .b8 __cudaparmf1__Z23UnpremultiplyComponents6float414IR_PixelFormat[16], .param .s32 __cudaparmf2__Z23UnpremultiplyComponents6float414IR_PixelFormat)
	{
	.reg .u32 %r<28>;
	.reg .f32 %f<38>;
	.reg .pred %p<19>;
	.loc	20	432	0
$LDWbegin__Z23UnpremultiplyComponents6float414IR_PixelFormat:
	ld.param.f32 	%f1, [__cudaparmf1__Z23UnpremultiplyComponents6float414IR_PixelFormat+0];
	mov.f32 	%f2, %f1;
	ld.param.f32 	%f3, [__cudaparmf1__Z23UnpremultiplyComponents6float414IR_PixelFormat+4];
	mov.f32 	%f4, %f3;
	ld.param.f32 	%f5, [__cudaparmf1__Z23UnpremultiplyComponents6float414IR_PixelFormat+8];
	mov.f32 	%f6, %f5;
	ld.param.f32 	%f7, [__cudaparmf1__Z23UnpremultiplyComponents6float414IR_PixelFormat+12];
	mov.f32 	%f8, %f7;
	ld.param.u32 	%r1, [__cudaparmf2__Z23UnpremultiplyComponents6float414IR_PixelFormat];
	mov.s32 	%r2, %r1;
	.loc	20	433	0
	mov.f32 	%f9, %f2;
	mov.f32 	%f10, %f4;
	mov.f32 	%f11, %f6;
	mov.f32 	%f12, %f8;
	.loc	20	435	0
	and.b32 	%r3, %r2, 448;
	mov.s32 	%r4, %r2;
	mov.s32 	%r5, 256;
	setp.ne.s32 	%p1, %r3, %r5;
	and.b32 	%r6, %r4, 1;
	selp.s32 	%r7, 1, 0, %p1;
	and.b32 	%r8, %r6, %r7;
	mov.u32 	%r9, 0;
	setp.eq.s32 	%p2, %r8, %r9;
	@%p2 bra 	$Lt_21_30210;
	.loc	20	57	0
	mov.u32 	%r10, 0;
	setp.ne.s32 	%p3, %r3, %r10;
	@%p3 bra 	$Lt_21_19458;
	.loc	20	59	0
	mov.f32 	%f13, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_198_9;
$Lt_21_19458:
	.loc	20	61	0
	mov.u32 	%r11, 64;
	setp.ne.s32 	%p4, %r3, %r11;
	@%p4 bra 	$Lt_21_19714;
	.loc	20	63	0
	mov.f32 	%f13, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_198_9;
$Lt_21_19714:
	.loc	20	65	0
	mov.u32 	%r12, 128;
	setp.ne.s32 	%p5, %r3, %r12;
	@%p5 bra 	$Lt_21_19970;
	.loc	20	68	0
	mov.f32 	%f13, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_198_9;
$Lt_21_19970:
	.loc	20	70	0
	mov.u32 	%r13, 192;
	setp.ne.s32 	%p6, %r3, %r13;
	@%p6 bra 	$Lt_21_20226;
	.loc	20	72	0
	mov.f32 	%f13, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_198_9;
$Lt_21_20226:
	.loc	20	76	0
	mov.f32 	%f13, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_198_9:
	.loc	20	118	0
	and.b32 	%r14, %r2, 2048;
	mov.s32 	%r15, 0;
	setp.ne.s32 	%p7, %r14, %r15;
	@!%p7 bra 	$Lt_21_30978;
	.loc	20	100	0
	ld.const.f32 	%f14, [kYCbCrOffset+0];
	bra.uni 	$Lt_21_30722;
$Lt_21_30978:
	ld.const.f32 	%f14, [kYCbCrFullRangeOffset+0];
$Lt_21_30722:
	.loc	20	118	0
	@!%p7 bra 	$Lt_21_31490;
	.loc	20	100	0
	ld.const.f32 	%f15, [kYCbCrOffset+4];
	bra.uni 	$Lt_21_31234;
$Lt_21_31490:
	ld.const.f32 	%f15, [kYCbCrFullRangeOffset+4];
$Lt_21_31234:
	.loc	20	118	0
	@!%p7 bra 	$Lt_21_32002;
	.loc	20	100	0
	ld.const.f32 	%f16, [kYCbCrOffset+8];
	bra.uni 	$Lt_21_31746;
$Lt_21_32002:
	ld.const.f32 	%f16, [kYCbCrFullRangeOffset+8];
$Lt_21_31746:
	.loc	20	437	0
	mov.f32 	%f17, 0f437f0000;    	// 255
	div.approx.ftz.f32 	%f18, %f13, %f17;
	mul.ftz.f32 	%f19, %f18, %f14;
	sub.ftz.f32 	%f10, %f4, %f19;
	mul.ftz.f32 	%f20, %f18, %f15;
	sub.ftz.f32 	%f11, %f6, %f20;
	mul.ftz.f32 	%f21, %f18, %f16;
	sub.ftz.f32 	%f12, %f8, %f21;
$Lt_21_30210:
	mov.f32 	%f22, 0fb70637bd;    	// -8e-006
	add.ftz.f32 	%f23, %f2, %f22;
	mov.f32 	%f24, 0f00000000;    	// 0
	setp.le.ftz.f32 	%p8, %f23, %f24;
	@!%p8 bra 	$Lt_21_32514;
	mov.f32 	%f12, 0f00000000;    	// 0
	mov.f32 	%f11, 0f00000000;    	// 0
	mov.f32 	%f10, 0f00000000;    	// 0
	mov.f32 	%f9, 0f00000000;     	// 0
	bra.uni 	$Lt_21_32258;
$Lt_21_32514:
	.loc	20	57	0
	mov.u32 	%r16, 0;
	setp.ne.s32 	%p9, %r3, %r16;
	@%p9 bra 	$Lt_21_22530;
	.loc	20	59	0
	mov.f32 	%f25, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_198_5;
$Lt_21_22530:
	.loc	20	61	0
	mov.u32 	%r17, 64;
	setp.ne.s32 	%p10, %r3, %r17;
	@%p10 bra 	$Lt_21_22786;
	.loc	20	63	0
	mov.f32 	%f25, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_198_5;
$Lt_21_22786:
	.loc	20	65	0
	mov.u32 	%r18, 128;
	setp.ne.s32 	%p11, %r3, %r18;
	@%p11 bra 	$Lt_21_23042;
	.loc	20	68	0
	mov.f32 	%f25, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_198_5;
$Lt_21_23042:
	.loc	20	70	0
	mov.u32 	%r19, 192;
	setp.ne.s32 	%p12, %r3, %r19;
	@%p12 bra 	$Lt_21_23298;
	.loc	20	72	0
	mov.f32 	%f25, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_198_5;
$Lt_21_23298:
	.loc	20	76	0
	mov.f32 	%f25, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_198_5:
	.loc	20	447	0
	div.approx.ftz.f32 	%f26, %f25, %f2;
	mul.ftz.f32 	%f10, %f26, %f10;
	.loc	20	448	0
	mul.ftz.f32 	%f11, %f26, %f11;
	.loc	20	449	0
	mul.ftz.f32 	%f12, %f26, %f12;
$Lt_21_32258:
	.loc	20	452	0
	mov.u32 	%r20, 0;
	setp.eq.s32 	%p13, %r8, %r20;
	@%p13 bra 	$Lt_21_32770;
	.loc	20	57	0
	mov.u32 	%r21, 0;
	setp.ne.s32 	%p14, %r3, %r21;
	@%p14 bra 	$Lt_21_23810;
	.loc	20	59	0
	mov.f32 	%f27, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_198_3;
$Lt_21_23810:
	.loc	20	61	0
	mov.u32 	%r22, 64;
	setp.ne.s32 	%p15, %r3, %r22;
	@%p15 bra 	$Lt_21_24066;
	.loc	20	63	0
	mov.f32 	%f27, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_198_3;
$Lt_21_24066:
	.loc	20	65	0
	mov.u32 	%r23, 128;
	setp.ne.s32 	%p16, %r3, %r23;
	@%p16 bra 	$Lt_21_24322;
	.loc	20	68	0
	mov.f32 	%f27, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_198_3;
$Lt_21_24322:
	.loc	20	70	0
	mov.u32 	%r24, 192;
	setp.ne.s32 	%p17, %r3, %r24;
	@%p17 bra 	$Lt_21_24578;
	.loc	20	72	0
	mov.f32 	%f27, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_198_3;
$Lt_21_24578:
	.loc	20	76	0
	mov.f32 	%f27, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_198_3:
	.loc	20	107	0
	and.b32 	%r25, %r2, 2048;
	mov.s32 	%r26, 0;
	setp.ne.s32 	%p7, %r25, %r26;
	@!%p7 bra 	$Lt_21_33538;
	.loc	20	100	0
	ld.const.f32 	%f28, [kYCbCrOffset+0];
	bra.uni 	$Lt_21_33282;
$Lt_21_33538:
	ld.const.f32 	%f28, [kYCbCrFullRangeOffset+0];
$Lt_21_33282:
	.loc	20	107	0
	@!%p7 bra 	$Lt_21_34050;
	.loc	20	100	0
	ld.const.f32 	%f29, [kYCbCrOffset+4];
	bra.uni 	$Lt_21_33794;
$Lt_21_34050:
	ld.const.f32 	%f29, [kYCbCrFullRangeOffset+4];
$Lt_21_33794:
	.loc	20	107	0
	@!%p7 bra 	$Lt_21_34562;
	.loc	20	100	0
	ld.const.f32 	%f30, [kYCbCrOffset+8];
	bra.uni 	$Lt_21_34306;
$Lt_21_34562:
	ld.const.f32 	%f30, [kYCbCrFullRangeOffset+8];
$Lt_21_34306:
	.loc	20	454	0
	mov.f32 	%f31, 0f437f0000;    	// 255
	div.approx.ftz.f32 	%f32, %f27, %f31;
	fma.rn.ftz.f32 	%f10, %f32, %f28, %f10;
	fma.rn.ftz.f32 	%f11, %f32, %f29, %f11;
	fma.rn.ftz.f32 	%f12, %f32, %f30, %f12;
$Lt_21_32770:
	.loc	20	457	0
	mov.f32 	%f33, %f9;
	st.param.f32 	[__cudaretf__Z23UnpremultiplyComponents6float414IR_PixelFormat+0], %f33;
	mov.f32 	%f34, %f10;
	st.param.f32 	[__cudaretf__Z23UnpremultiplyComponents6float414IR_PixelFormat+4], %f34;
	mov.f32 	%f35, %f11;
	st.param.f32 	[__cudaretf__Z23UnpremultiplyComponents6float414IR_PixelFormat+8], %f35;
	mov.f32 	%f36, %f12;
	st.param.f32 	[__cudaretf__Z23UnpremultiplyComponents6float414IR_PixelFormat+12], %f36;
	ret;
$LDWend__Z23UnpremultiplyComponents6float414IR_PixelFormat:
	} // _Z23UnpremultiplyComponents6float414IR_PixelFormat

	.visible .func (.param .align 16 .b8 __cudaretf__Z23ConvertPixel_444_To_4446float414IR_PixelFormatS0_[16]) _Z23ConvertPixel_444_To_4446float414IR_PixelFormatS0_ (.param .align 16 .b8 __cudaparmf1__Z23ConvertPixel_444_To_4446float414IR_PixelFormatS0_[16], .param .s32 __cudaparmf2__Z23ConvertPixel_444_To_4446float414IR_PixelFormatS0_, .param .s32 __cudaparmf3__Z23ConvertPixel_444_To_4446float414IR_PixelFormatS0_)
	{
	.reg .u32 %r<185>;
	.reg .u64 %rd<3>;
	.reg .f32 %f<198>;
	.reg .pred %p<135>;
	.loc	20	468	0
$LDWbegin__Z23ConvertPixel_444_To_4446float414IR_PixelFormatS0_:
	ld.param.f32 	%f1, [__cudaparmf1__Z23ConvertPixel_444_To_4446float414IR_PixelFormatS0_+0];
	mov.f32 	%f2, %f1;
	ld.param.f32 	%f3, [__cudaparmf1__Z23ConvertPixel_444_To_4446float414IR_PixelFormatS0_+4];
	mov.f32 	%f4, %f3;
	ld.param.f32 	%f5, [__cudaparmf1__Z23ConvertPixel_444_To_4446float414IR_PixelFormatS0_+8];
	mov.f32 	%f6, %f5;
	ld.param.f32 	%f7, [__cudaparmf1__Z23ConvertPixel_444_To_4446float414IR_PixelFormatS0_+12];
	mov.f32 	%f8, %f7;
	ld.param.u32 	%r1, [__cudaparmf2__Z23ConvertPixel_444_To_4446float414IR_PixelFormatS0_];
	mov.s32 	%r2, %r1;
	ld.param.u32 	%r3, [__cudaparmf3__Z23ConvertPixel_444_To_4446float414IR_PixelFormatS0_];
	mov.s32 	%r4, %r3;
	.loc	20	469	0
	mov.f32 	%f9, %f2;
	mov.f32 	%f10, %f4;
	mov.f32 	%f11, %f6;
	mov.f32 	%f12, %f8;
	and.b32 	%r5, %r2, 4096;
	mov.u32 	%r6, 0;
	setp.ne.s32 	%p1, %r5, %r6;
	@%p1 bra 	$Lt_22_222722;
	.loc	20	473	0
	mov.f32 	%f9, %f8;
	mov.f32 	%f10, %f6;
	mov.f32 	%f11, %f4;
	mov.f32 	%f12, %f2;
$Lt_22_222722:
	.loc	20	476	0
	and.b32 	%r7, %r2, 448;
	mov.s32 	%r8, %r2;
	and.b32 	%r9, %r4, 448;
	mov.s32 	%r10, %r4;
	mov.s32 	%r11, 256;
	setp.ne.s32 	%p2, %r7, %r11;
	and.b32 	%r12, %r8, 1;
	mov.s32 	%r13, 256;
	setp.ne.s32 	%p3, %r9, %r13;
	and.b32 	%r14, %r10, 1;
	selp.s32 	%r15, 1, 0, %p2;
	selp.s32 	%r16, 1, 0, %p3;
	and.b32 	%r17, %r12, %r15;
	and.b32 	%r18, %r14, %r16;
	mov.u32 	%r19, 0;
	setp.eq.s32 	%p4, %r17, %r19;
	@%p4 bra 	$Lt_22_249858;
	mov.u32 	%r20, 0;
	setp.ne.s32 	%p5, %r18, %r20;
	@%p5 bra 	$Lt_22_249858;
	.loc	20	57	0
	mov.u32 	%r21, 0;
	setp.ne.s32 	%p6, %r7, %r21;
	@%p6 bra 	$Lt_22_140802;
	.loc	20	59	0
	mov.f32 	%f13, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_75;
$Lt_22_140802:
	.loc	20	61	0
	mov.u32 	%r22, 64;
	setp.ne.s32 	%p7, %r7, %r22;
	@%p7 bra 	$Lt_22_141058;
	.loc	20	63	0
	mov.f32 	%f13, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_75;
$Lt_22_141058:
	.loc	20	65	0
	mov.u32 	%r23, 128;
	setp.ne.s32 	%p8, %r7, %r23;
	@%p8 bra 	$Lt_22_141314;
	.loc	20	68	0
	mov.f32 	%f13, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_75;
$Lt_22_141314:
	.loc	20	70	0
	mov.u32 	%r24, 192;
	setp.ne.s32 	%p9, %r7, %r24;
	@%p9 bra 	$Lt_22_141570;
	.loc	20	72	0
	mov.f32 	%f13, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_75;
$Lt_22_141570:
	.loc	20	76	0
	mov.f32 	%f13, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_199_75:
	.loc	20	118	0
	and.b32 	%r25, %r2, 2048;
	mov.s32 	%r26, 0;
	setp.ne.s32 	%p10, %r25, %r26;
	@!%p10 bra 	$Lt_22_223490;
	.loc	20	100	0
	ld.const.f32 	%f14, [kYCbCrOffset+0];
	bra.uni 	$Lt_22_223234;
$Lt_22_223490:
	ld.const.f32 	%f14, [kYCbCrFullRangeOffset+0];
$Lt_22_223234:
	.loc	20	118	0
	@!%p10 bra 	$Lt_22_224002;
	.loc	20	100	0
	ld.const.f32 	%f15, [kYCbCrOffset+4];
	bra.uni 	$Lt_22_223746;
$Lt_22_224002:
	ld.const.f32 	%f15, [kYCbCrFullRangeOffset+4];
$Lt_22_223746:
	.loc	20	118	0
	@!%p10 bra 	$Lt_22_224514;
	.loc	20	100	0
	ld.const.f32 	%f16, [kYCbCrOffset+8];
	bra.uni 	$Lt_22_224258;
$Lt_22_224514:
	ld.const.f32 	%f16, [kYCbCrFullRangeOffset+8];
$Lt_22_224258:
	.loc	20	478	0
	mov.f32 	%f17, 0f437f0000;    	// 255
	div.approx.ftz.f32 	%f18, %f13, %f17;
	mul.ftz.f32 	%f19, %f18, %f14;
	sub.ftz.f32 	%f10, %f10, %f19;
	mul.ftz.f32 	%f20, %f18, %f15;
	sub.ftz.f32 	%f11, %f11, %f20;
	mul.ftz.f32 	%f21, %f18, %f16;
	sub.ftz.f32 	%f12, %f12, %f21;
$Lt_22_249858:
$Lt_22_26114:
	.loc	20	481	0
	and.b32 	%r27, %r2, 2;
	and.b32 	%r28, %r4, 2;
	mov.u32 	%r29, 0;
	setp.eq.s32 	%p11, %r27, %r29;
	@%p11 bra 	$Lt_22_250370;
	mov.u32 	%r30, 0;
	setp.ne.s32 	%p12, %r28, %r30;
	@%p12 bra 	$Lt_22_250370;
	.loc	20	483	0
	mov.f32 	%f22, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p13, %f10, %f22;
	@!%p13 bra 	$Lt_22_224770;
	.loc	20	372	0
	neg.ftz.f32 	%f23, %f10;
	lg2.approx.ftz.f32 	%f24, %f23;
	mov.f32 	%f25, 0f3ee66666;    	// 0.45
	mul.ftz.f32 	%f26, %f24, %f25;
	ex2.approx.ftz.f32 	%f27, %f26;
	neg.ftz.f32 	%f28, %f27;
	bra.uni 	$LDWendi___log2f_199_71;
$Lt_22_224770:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f29, %f10;
	mov.f32 	%f30, 0f3ee66666;    	// 0.45
	mul.ftz.f32 	%f31, %f29, %f30;
	ex2.approx.ftz.f32 	%f28, %f31;
$LDWendi___log2f_199_71:
	.loc	20	483	0
	mov.f32 	%f32, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p14, %f11, %f32;
	@!%p14 bra 	$Lt_22_225282;
	.loc	20	372	0
	neg.ftz.f32 	%f33, %f11;
	lg2.approx.ftz.f32 	%f34, %f33;
	mov.f32 	%f35, 0f3ee66666;    	// 0.45
	mul.ftz.f32 	%f36, %f34, %f35;
	ex2.approx.ftz.f32 	%f37, %f36;
	neg.ftz.f32 	%f38, %f37;
	bra.uni 	$LDWendi___log2f_199_69;
$Lt_22_225282:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f39, %f11;
	mov.f32 	%f40, 0f3ee66666;    	// 0.45
	mul.ftz.f32 	%f41, %f39, %f40;
	ex2.approx.ftz.f32 	%f38, %f41;
$LDWendi___log2f_199_69:
	.loc	20	483	0
	mov.f32 	%f42, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p15, %f12, %f42;
	@!%p15 bra 	$Lt_22_225794;
	.loc	20	372	0
	neg.ftz.f32 	%f43, %f12;
	lg2.approx.ftz.f32 	%f44, %f43;
	mov.f32 	%f45, 0f3ee66666;    	// 0.45
	mul.ftz.f32 	%f46, %f44, %f45;
	ex2.approx.ftz.f32 	%f47, %f46;
	neg.ftz.f32 	%f48, %f47;
	bra.uni 	$LDWendi___log2f_199_67;
$Lt_22_225794:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f49, %f12;
	mov.f32 	%f50, 0f3ee66666;    	// 0.45
	mul.ftz.f32 	%f51, %f49, %f50;
	ex2.approx.ftz.f32 	%f48, %f51;
$LDWendi___log2f_199_67:
	.loc	20	483	0
	mov.f32 	%f10, %f28;
	mov.f32 	%f11, %f38;
	mov.f32 	%f12, %f48;
$Lt_22_250370:
$Lt_22_29954:
	.loc	20	486	0
	and.b32 	%r31, %r2, 1;
	and.b32 	%r32, %r4, 1;
	and.b32 	%r33, %r2, 1536;
	and.b32 	%r34, %r4, 1536;
	set.ne.u32.s32 	%r35, %r31, %r32;
	neg.s32 	%r36, %r35;
	set.ne.u32.s32 	%r37, %r33, %r34;
	neg.s32 	%r38, %r37;
	or.b32 	%r39, %r36, %r38;
	mov.u32 	%r40, 0;
	setp.ne.s32 	%p16, %r39, %r40;
	@%p16 bra 	$Lt_22_80642;
	setp.eq.s32 	%p17, %r17, %r18;
	@%p17 bra 	$Lt_22_80898;
$Lt_22_80642:
	.loc	20	490	0
	mov.u32 	%r41, 0;
	setp.ne.s32 	%p18, %r31, %r41;
	@%p18 bra 	$Lt_22_226562;
	mov.s32 	%r42, 256;
	setp.eq.s32 	%p19, %r9, %r42;
	mov.u32 	%r43, 256;
	setp.ne.s32 	%p20, %r7, %r43;
	@%p20 bra 	$Lt_22_227074;
	.loc	20	137	0
	mov.s32 	%r44, 512;
	setp.eq.s32 	%p21, %r34, %r44;
	@!%p19 bra 	$Lt_22_144898;
	.loc	20	139	0
	@!%p21 bra 	$Lt_22_145154;
	.loc	20	141	0
	mov.u64 	%rd1, kRGB32f_To_709YPbPr;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__199_65;
$Lt_22_145154:
	.loc	20	145	0
	mov.u64 	%rd1, kRGB32f_To_601YPbPr;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__199_65;
$Lt_22_144898:
	.loc	20	150	0
	@!%p21 bra 	$Lt_22_145410;
	.loc	20	152	0
	mov.u64 	%rd1, kRGB32f_To_709YCbCr;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__199_65;
$Lt_22_145410:
	.loc	20	154	0
	and.b32 	%r45, %r4, 2048;
	mov.u32 	%r46, 0;
	setp.ne.s32 	%p22, %r45, %r46;
	@%p22 bra 	$Lt_22_145666;
	.loc	20	156	0
	mov.u64 	%rd1, kRGB32f_To_601YCbCrFullRange;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__199_65;
$Lt_22_145666:
	.loc	20	160	0
	mov.u64 	%rd1, kRGB32f_To_601YCbCr;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__199_65;
$Lt_22_227074:
	@!%p19 bra 	$Lt_22_227586;
	bra.uni 	$Lt_22_226306;
$Lt_22_227586:
	.loc	20	179	0
	mov.u32 	%r47, 512;
	setp.ne.s32 	%p23, %r34, %r47;
	@%p23 bra 	$Lt_22_146434;
	.loc	20	181	0
	mov.u64 	%rd1, kRGB8u_To_709YCbCr;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__199_65;
$Lt_22_146434:
	.loc	20	183	0
	and.b32 	%r48, %r4, 2048;
	mov.u32 	%r49, 0;
	setp.ne.s32 	%p24, %r48, %r49;
	@%p24 bra 	$Lt_22_146690;
	.loc	20	185	0
	mov.u64 	%rd1, kRGB8u_To_601YCbCrFullRange;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__199_65;
$Lt_22_146690:
	.loc	20	189	0
	mov.u64 	%rd1, kRGB8u_To_601YCbCr;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__199_65;
$Lt_22_226562:
	mov.s32 	%r50, 0;
	setp.eq.s32 	%p25, %r32, %r50;
	mov.u32 	%r51, 512;
	setp.ne.s32 	%p26, %r33, %r51;
	@%p26 bra 	$Lt_22_228098;
	mov.s32 	%r52, 256;
	setp.eq.s32 	%p27, %r7, %r52;
	@!%p25 bra 	$Lt_22_228610;
	mov.s32 	%r53, 256;
	setp.eq.s32 	%p28, %r9, %r53;
	@!%p27 bra 	$Lt_22_229122;
	@!%p28 bra 	$Lt_22_226306;
	.loc	20	202	0
	mov.u64 	%rd1, k709YPbPr_To_RGB32f;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__199_65;
$Lt_22_229122:
	.loc	20	211	0
	@!%p28 bra 	$Lt_22_147970;
	.loc	20	213	0
	mov.u64 	%rd1, k709YCbCr_To_RGB32f;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__199_65;
$Lt_22_147970:
	.loc	20	217	0
	mov.u64 	%rd1, k709YCbCr_To_RGB8u;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__199_65;
$Lt_22_228610:
	@!%p27 bra 	$Lt_22_230146;
	bra.uni 	$Lt_22_226306;
$Lt_22_230146:
	mov.s32 	%r54, 256;
	set.eq.u32.s32 	%r55, %r9, %r54;
	neg.s32 	%r56, %r55;
	and.b32 	%r57, %r4, 2048;
	mov.s32 	%r58, 0;
	set.eq.u32.s32 	%r59, %r57, %r58;
	neg.s32 	%r60, %r59;
	or.b32 	%r61, %r56, %r60;
	mov.u32 	%r62, 0;
	setp.eq.s32 	%p29, %r61, %r62;
	@%p29 bra 	$Lt_22_230658;
	bra.uni 	$Lt_22_226306;
$Lt_22_230658:
	.loc	20	250	0
	mov.u64 	%rd1, k709YCbCr_To_601YCbCr;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__199_65;
$Lt_22_228098:
	and.b32 	%r63, %r2, 2048;
	mov.s32 	%r64, 0;
	setp.eq.s32 	%p30, %r63, %r64;
	@!%p30 bra 	$Lt_22_231170;
	@!%p25 bra 	$Lt_22_226306;
	.loc	20	259	0
	mov.u32 	%r65, 256;
	setp.ne.s32 	%p31, %r9, %r65;
	@%p31 bra 	$Lt_22_149506;
	.loc	20	261	0
	mov.u64 	%rd1, k601YCbCrFullRange_To_RGB32f;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__199_65;
$Lt_22_149506:
	.loc	20	265	0
	mov.u64 	%rd1, k601YCbCrFullRange_To_RGB8u;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__199_65;
$Lt_22_231170:
	mov.s32 	%r66, 256;
	setp.eq.s32 	%p27, %r7, %r66;
	@!%p25 bra 	$Lt_22_232194;
	mov.s32 	%r67, 256;
	setp.eq.s32 	%p32, %r9, %r67;
	@!%p27 bra 	$Lt_22_232706;
	@!%p32 bra 	$Lt_22_226306;
	.loc	20	302	0
	mov.u64 	%rd1, k601YPbPr_To_RGB32f;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__199_65;
$Lt_22_232706:
	.loc	20	311	0
	@!%p32 bra 	$Lt_22_151298;
	.loc	20	313	0
	mov.u64 	%rd1, k601YCbCr_To_RGB32f;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__199_65;
$Lt_22_151298:
	.loc	20	317	0
	mov.u64 	%rd1, k601YCbCr_To_RGB8u;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__199_65;
$Lt_22_232194:
	@!%p27 bra 	$Lt_22_233730;
	bra.uni 	$Lt_22_226306;
$Lt_22_233730:
	selp.s32 	%r68, 1, 0, %p30;
	mov.s32 	%r69, 256;
	set.eq.u32.s32 	%r70, %r9, %r69;
	neg.s32 	%r71, %r70;
	or.b32 	%r72, %r68, %r71;
	mov.u32 	%r73, 0;
	setp.eq.s32 	%p33, %r72, %r73;
	@%p33 bra 	$Lt_22_234242;
	bra.uni 	$Lt_22_226306;
$Lt_22_234242:
	.loc	20	350	0
	mov.u64 	%rd1, k601YCbCr_To_709YCbCr;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__199_65;
$Lt_22_226306:
	.loc	20	355	0
	mov.u64 	%rd1, 0;
$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__199_65:
	.loc	20	490	0
	ld.global.f32 	%f52, [%rd1+16];
	mul.ftz.f32 	%f53, %f52, %f11;
	ld.global.f32 	%f54, [%rd1+12];
	fma.rn.ftz.f32 	%f55, %f54, %f10, %f53;
	ld.global.f32 	%f56, [%rd1+20];
	fma.rn.ftz.f32 	%f57, %f56, %f12, %f55;
	ld.global.f32 	%f58, [%rd1+28];
	mul.ftz.f32 	%f59, %f58, %f11;
	ld.global.f32 	%f60, [%rd1+24];
	fma.rn.ftz.f32 	%f61, %f60, %f10, %f59;
	ld.global.f32 	%f62, [%rd1+32];
	fma.rn.ftz.f32 	%f63, %f62, %f12, %f61;
	ld.global.f32 	%f64, [%rd1+4];
	mul.ftz.f32 	%f65, %f64, %f11;
	ld.global.f32 	%f66, [%rd1+0];
	fma.rn.ftz.f32 	%f67, %f66, %f10, %f65;
	ld.global.f32 	%f68, [%rd1+8];
	fma.rn.ftz.f32 	%f10, %f68, %f12, %f67;
	mov.f32 	%f11, %f57;
	mov.f32 	%f12, %f63;
	setp.eq.s32 	%p34, %r7, %r9;
	@%p34 bra 	$Lt_22_235010;
	.loc	20	494	0
	mov.s32 	%r74, 256;
	setp.eq.s32 	%p27, %r7, %r74;
	@!%p27 bra 	$L_22_220162;
	mov.s32 	%r75, 0;
	setp.eq.s32 	%p35, %r9, %r75;
	@%p35 bra 	$Lt_22_251394;
$L_22_220162:
	mov.s32 	%r76, 0;
	setp.eq.s32 	%p36, %r7, %r76;
	@!%p36 bra 	$Lt_22_251650;
	mov.u32 	%r77, 256;
	setp.ne.s32 	%p37, %r9, %r77;
	@%p37 bra 	$Lt_22_251650;
	mov.s32 	%r78, 0;
	setp.eq.s32 	%p35, %r9, %r78;
	bra.uni 	$L_22_219906;
$Lt_22_251394:
	mov.s32 	%r79, 0;
	setp.eq.s32 	%p36, %r7, %r79;
$L_22_219906:
	.loc	20	57	0
	@!%p35 bra 	$Lt_22_152834;
	.loc	20	59	0
	mov.f32 	%f69, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_63;
$Lt_22_152834:
	.loc	20	61	0
	mov.u32 	%r80, 64;
	setp.ne.s32 	%p38, %r9, %r80;
	@%p38 bra 	$Lt_22_153090;
	.loc	20	63	0
	mov.f32 	%f69, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_63;
$Lt_22_153090:
	.loc	20	65	0
	mov.u32 	%r81, 128;
	setp.ne.s32 	%p39, %r9, %r81;
	@%p39 bra 	$Lt_22_153346;
	.loc	20	68	0
	mov.f32 	%f69, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_63;
$Lt_22_153346:
	.loc	20	70	0
	mov.u32 	%r82, 192;
	setp.ne.s32 	%p40, %r9, %r82;
	@%p40 bra 	$Lt_22_153602;
	.loc	20	72	0
	mov.f32 	%f69, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_63;
$Lt_22_153602:
	.loc	20	76	0
	mov.f32 	%f69, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_199_63:
	.loc	20	57	0
	@!%p36 bra 	$Lt_22_153858;
	.loc	20	59	0
	mov.f32 	%f70, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_61;
$Lt_22_153858:
	.loc	20	61	0
	mov.u32 	%r83, 64;
	setp.ne.s32 	%p41, %r7, %r83;
	@%p41 bra 	$Lt_22_154114;
	.loc	20	63	0
	mov.f32 	%f70, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_61;
$Lt_22_154114:
	.loc	20	65	0
	mov.u32 	%r84, 128;
	setp.ne.s32 	%p42, %r7, %r84;
	@%p42 bra 	$Lt_22_154370;
	.loc	20	68	0
	mov.f32 	%f70, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_61;
$Lt_22_154370:
	.loc	20	70	0
	mov.u32 	%r85, 192;
	setp.ne.s32 	%p43, %r7, %r85;
	@%p43 bra 	$Lt_22_154626;
	.loc	20	72	0
	mov.f32 	%f70, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_61;
$Lt_22_154626:
	.loc	20	76	0
	mov.f32 	%f70, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_199_61:
	.loc	20	498	0
	div.approx.ftz.f32 	%f71, %f69, %f70;
	mul.ftz.f32 	%f9, %f9, %f71;
	bra.uni 	$Lt_22_235010;
$Lt_22_251650:
$L_22_219650:
	.loc	20	500	0
	@!%p27 bra 	$L_22_221186;
	@%p3 bra 	$L_22_220930;
$L_22_221186:
	@!%p2 bra 	$Lt_22_252674;
	mov.u32 	%r86, 256;
	setp.ne.s32 	%p44, %r9, %r86;
	@%p44 bra 	$Lt_22_252674;
$L_22_220930:
	.loc	20	57	0
	mov.u32 	%r87, 0;
	setp.ne.s32 	%p45, %r9, %r87;
	@%p45 bra 	$Lt_22_155138;
	.loc	20	59	0
	mov.f32 	%f69, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_59;
$Lt_22_155138:
	.loc	20	61	0
	mov.u32 	%r88, 64;
	setp.ne.s32 	%p46, %r9, %r88;
	@%p46 bra 	$Lt_22_155394;
	.loc	20	63	0
	mov.f32 	%f69, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_59;
$Lt_22_155394:
	.loc	20	65	0
	mov.u32 	%r89, 128;
	setp.ne.s32 	%p47, %r9, %r89;
	@%p47 bra 	$Lt_22_155650;
	.loc	20	68	0
	mov.f32 	%f69, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_59;
$Lt_22_155650:
	.loc	20	70	0
	mov.u32 	%r90, 192;
	setp.ne.s32 	%p48, %r9, %r90;
	@%p48 bra 	$Lt_22_155906;
	.loc	20	72	0
	mov.f32 	%f69, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_59;
$Lt_22_155906:
	.loc	20	76	0
	mov.f32 	%f69, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_199_59:
	.loc	20	57	0
	@!%p36 bra 	$Lt_22_156162;
	.loc	20	59	0
	mov.f32 	%f70, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_57;
$Lt_22_156162:
	.loc	20	61	0
	mov.u32 	%r91, 64;
	setp.ne.s32 	%p49, %r7, %r91;
	@%p49 bra 	$Lt_22_156418;
	.loc	20	63	0
	mov.f32 	%f70, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_57;
$Lt_22_156418:
	.loc	20	65	0
	mov.u32 	%r92, 128;
	setp.ne.s32 	%p50, %r7, %r92;
	@%p50 bra 	$Lt_22_156674;
	.loc	20	68	0
	mov.f32 	%f70, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_57;
$Lt_22_156674:
	.loc	20	70	0
	mov.u32 	%r93, 192;
	setp.ne.s32 	%p51, %r7, %r93;
	@%p51 bra 	$Lt_22_156930;
	.loc	20	72	0
	mov.f32 	%f70, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_57;
$Lt_22_156930:
	.loc	20	76	0
	mov.f32 	%f70, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_199_57:
	.loc	20	504	0
	div.approx.ftz.f32 	%f72, %f69, %f70;
	mul.ftz.f32 	%f9, %f9, %f72;
	.loc	20	57	0
	@!%p36 bra 	$Lt_22_158210;
	.loc	20	59	0
	mov.f32 	%f70, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_53;
$Lt_22_158210:
	.loc	20	61	0
	mov.u32 	%r94, 64;
	setp.ne.s32 	%p52, %r7, %r94;
	@%p52 bra 	$Lt_22_158466;
	.loc	20	63	0
	mov.f32 	%f70, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_53;
$Lt_22_158466:
	.loc	20	65	0
	mov.u32 	%r95, 128;
	setp.ne.s32 	%p53, %r7, %r95;
	@%p53 bra 	$Lt_22_158722;
	.loc	20	68	0
	mov.f32 	%f70, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_53;
$Lt_22_158722:
	.loc	20	70	0
	mov.u32 	%r96, 192;
	setp.ne.s32 	%p54, %r7, %r96;
	@%p54 bra 	$Lt_22_158978;
	.loc	20	72	0
	mov.f32 	%f70, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_53;
$Lt_22_158978:
	.loc	20	76	0
	mov.f32 	%f70, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_199_53:
	.loc	20	505	0
	mov.f32 	%f73, 0f437f0000;    	// 255
	div.approx.ftz.f32 	%f74, %f73, %f70;
	mul.ftz.f32 	%f10, %f10, %f74;
	.loc	20	57	0
	@!%p36 bra 	$Lt_22_160258;
	.loc	20	59	0
	mov.f32 	%f70, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_49;
$Lt_22_160258:
	.loc	20	61	0
	mov.u32 	%r97, 64;
	setp.ne.s32 	%p55, %r7, %r97;
	@%p55 bra 	$Lt_22_160514;
	.loc	20	63	0
	mov.f32 	%f70, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_49;
$Lt_22_160514:
	.loc	20	65	0
	mov.u32 	%r98, 128;
	setp.ne.s32 	%p56, %r7, %r98;
	@%p56 bra 	$Lt_22_160770;
	.loc	20	68	0
	mov.f32 	%f70, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_49;
$Lt_22_160770:
	.loc	20	70	0
	mov.u32 	%r99, 192;
	setp.ne.s32 	%p57, %r7, %r99;
	@%p57 bra 	$Lt_22_161026;
	.loc	20	72	0
	mov.f32 	%f70, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_49;
$Lt_22_161026:
	.loc	20	76	0
	mov.f32 	%f70, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_199_49:
	.loc	20	506	0
	mov.f32 	%f75, 0f437f0000;    	// 255
	div.approx.ftz.f32 	%f76, %f75, %f70;
	mul.ftz.f32 	%f11, %f57, %f76;
	.loc	20	57	0
	@!%p36 bra 	$Lt_22_162306;
	.loc	20	59	0
	mov.f32 	%f70, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_45;
$Lt_22_162306:
	.loc	20	61	0
	mov.u32 	%r100, 64;
	setp.ne.s32 	%p58, %r7, %r100;
	@%p58 bra 	$Lt_22_162562;
	.loc	20	63	0
	mov.f32 	%f70, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_45;
$Lt_22_162562:
	.loc	20	65	0
	mov.u32 	%r101, 128;
	setp.ne.s32 	%p59, %r7, %r101;
	@%p59 bra 	$Lt_22_162818;
	.loc	20	68	0
	mov.f32 	%f70, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_45;
$Lt_22_162818:
	.loc	20	70	0
	mov.u32 	%r102, 192;
	setp.ne.s32 	%p60, %r7, %r102;
	@%p60 bra 	$Lt_22_163074;
	.loc	20	72	0
	mov.f32 	%f70, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_45;
$Lt_22_163074:
	.loc	20	76	0
	mov.f32 	%f70, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_199_45:
	.loc	20	507	0
	mov.f32 	%f77, 0f437f0000;    	// 255
	div.approx.ftz.f32 	%f78, %f77, %f70;
	mul.ftz.f32 	%f12, %f63, %f78;
	bra.uni 	$Lt_22_235010;
$Lt_22_252674:
$L_22_220674:
	.loc	20	57	0
	@!%p36 bra 	$Lt_22_163330;
	.loc	20	59	0
	mov.f32 	%f69, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_43;
$Lt_22_163330:
	.loc	20	61	0
	mov.u32 	%r103, 64;
	setp.ne.s32 	%p61, %r7, %r103;
	@%p61 bra 	$Lt_22_163586;
	.loc	20	63	0
	mov.f32 	%f69, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_43;
$Lt_22_163586:
	.loc	20	65	0
	mov.u32 	%r104, 128;
	setp.ne.s32 	%p62, %r7, %r104;
	@%p62 bra 	$Lt_22_163842;
	.loc	20	68	0
	mov.f32 	%f69, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_43;
$Lt_22_163842:
	.loc	20	70	0
	mov.u32 	%r105, 192;
	setp.ne.s32 	%p63, %r7, %r105;
	@%p63 bra 	$Lt_22_164098;
	.loc	20	72	0
	mov.f32 	%f69, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_43;
$Lt_22_164098:
	.loc	20	76	0
	mov.f32 	%f69, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_199_43:
	.loc	20	511	0
	mov.f32 	%f79, 0f437f0000;    	// 255
	div.approx.ftz.f32 	%f80, %f69, %f79;
	mul.ftz.f32 	%f9, %f80, %f9;
	mul.ftz.f32 	%f10, %f80, %f10;
	mul.ftz.f32 	%f11, %f80, %f57;
	mul.ftz.f32 	%f12, %f80, %f63;
	bra.uni 	$Lt_22_235010;
$Lt_22_80898:
	.loc	20	486	0
	setp.eq.s32 	%p64, %r7, %r9;
	@%p64 bra 	$Lt_22_235010;
	.loc	20	57	0
	mov.u32 	%r106, 0;
	setp.ne.s32 	%p65, %r9, %r106;
	@%p65 bra 	$Lt_22_165634;
	.loc	20	59	0
	mov.f32 	%f69, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_39;
$Lt_22_165634:
	.loc	20	61	0
	mov.u32 	%r107, 64;
	setp.ne.s32 	%p66, %r9, %r107;
	@%p66 bra 	$Lt_22_165890;
	.loc	20	63	0
	mov.f32 	%f69, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_39;
$Lt_22_165890:
	.loc	20	65	0
	mov.u32 	%r108, 128;
	setp.ne.s32 	%p67, %r9, %r108;
	@%p67 bra 	$Lt_22_166146;
	.loc	20	68	0
	mov.f32 	%f69, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_39;
$Lt_22_166146:
	.loc	20	70	0
	mov.u32 	%r109, 192;
	setp.ne.s32 	%p68, %r9, %r109;
	@%p68 bra 	$Lt_22_166402;
	.loc	20	72	0
	mov.f32 	%f69, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_39;
$Lt_22_166402:
	.loc	20	76	0
	mov.f32 	%f69, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_199_39:
	.loc	20	57	0
	mov.u32 	%r110, 0;
	setp.ne.s32 	%p69, %r7, %r110;
	@%p69 bra 	$Lt_22_166658;
	.loc	20	59	0
	mov.f32 	%f70, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_37;
$Lt_22_166658:
	.loc	20	61	0
	mov.u32 	%r111, 64;
	setp.ne.s32 	%p70, %r7, %r111;
	@%p70 bra 	$Lt_22_166914;
	.loc	20	63	0
	mov.f32 	%f70, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_37;
$Lt_22_166914:
	.loc	20	65	0
	mov.u32 	%r112, 128;
	setp.ne.s32 	%p71, %r7, %r112;
	@%p71 bra 	$Lt_22_167170;
	.loc	20	68	0
	mov.f32 	%f70, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_37;
$Lt_22_167170:
	.loc	20	70	0
	mov.u32 	%r113, 192;
	setp.ne.s32 	%p72, %r7, %r113;
	@%p72 bra 	$Lt_22_167426;
	.loc	20	72	0
	mov.f32 	%f70, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_37;
$Lt_22_167426:
	.loc	20	76	0
	mov.f32 	%f70, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_199_37:
	.loc	20	517	0
	div.approx.ftz.f32 	%f81, %f69, %f70;
	mul.ftz.f32 	%f9, %f81, %f9;
	mul.ftz.f32 	%f10, %f81, %f10;
	mul.ftz.f32 	%f11, %f81, %f11;
	mul.ftz.f32 	%f12, %f81, %f12;
$Lt_22_235010:
$Lt_22_83202:
	.loc	20	520	0
	mov.u32 	%r114, 0;
	setp.eq.s32 	%p73, %r28, %r114;
	@%p73 bra 	$Lt_22_253186;
	mov.u32 	%r115, 0;
	setp.ne.s32 	%p74, %r27, %r115;
	@%p74 bra 	$Lt_22_253186;
	.loc	20	522	0
	mov.f32 	%f82, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p75, %f10, %f82;
	@!%p75 bra 	$Lt_22_235522;
	.loc	20	372	0
	neg.ftz.f32 	%f83, %f10;
	lg2.approx.ftz.f32 	%f84, %f83;
	mov.f32 	%f85, 0f400e38e4;    	// 2.22222
	mul.ftz.f32 	%f86, %f84, %f85;
	ex2.approx.ftz.f32 	%f87, %f86;
	neg.ftz.f32 	%f88, %f87;
	bra.uni 	$LDWendi___log2f_199_35;
$Lt_22_235522:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f89, %f10;
	mov.f32 	%f90, 0f400e38e4;    	// 2.22222
	mul.ftz.f32 	%f91, %f89, %f90;
	ex2.approx.ftz.f32 	%f88, %f91;
$LDWendi___log2f_199_35:
	.loc	20	522	0
	mov.f32 	%f92, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p76, %f11, %f92;
	@!%p76 bra 	$Lt_22_236034;
	.loc	20	372	0
	neg.ftz.f32 	%f93, %f11;
	lg2.approx.ftz.f32 	%f94, %f93;
	mov.f32 	%f95, 0f400e38e4;    	// 2.22222
	mul.ftz.f32 	%f96, %f94, %f95;
	ex2.approx.ftz.f32 	%f97, %f96;
	neg.ftz.f32 	%f98, %f97;
	bra.uni 	$LDWendi___log2f_199_33;
$Lt_22_236034:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f99, %f11;
	mov.f32 	%f100, 0f400e38e4;   	// 2.22222
	mul.ftz.f32 	%f101, %f99, %f100;
	ex2.approx.ftz.f32 	%f98, %f101;
$LDWendi___log2f_199_33:
	.loc	20	522	0
	mov.f32 	%f102, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p77, %f12, %f102;
	@!%p77 bra 	$Lt_22_236546;
	.loc	20	372	0
	neg.ftz.f32 	%f103, %f12;
	lg2.approx.ftz.f32 	%f104, %f103;
	mov.f32 	%f105, 0f400e38e4;   	// 2.22222
	mul.ftz.f32 	%f106, %f104, %f105;
	ex2.approx.ftz.f32 	%f107, %f106;
	neg.ftz.f32 	%f108, %f107;
	bra.uni 	$LDWendi___log2f_199_31;
$Lt_22_236546:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f109, %f12;
	mov.f32 	%f110, 0f400e38e4;   	// 2.22222
	mul.ftz.f32 	%f111, %f109, %f110;
	ex2.approx.ftz.f32 	%f108, %f111;
$LDWendi___log2f_199_31:
	.loc	20	522	0
	mov.f32 	%f10, %f88;
	mov.f32 	%f11, %f98;
	mov.f32 	%f12, %f108;
$Lt_22_253186:
$Lt_22_85250:
	.loc	20	525	0
	mov.u32 	%r116, 0;
	setp.eq.s32 	%p78, %r18, %r116;
	@%p78 bra 	$Lt_22_253698;
	mov.u32 	%r117, 0;
	setp.ne.s32 	%p79, %r17, %r117;
	@%p79 bra 	$Lt_22_253698;
	.loc	20	57	0
	mov.u32 	%r118, 0;
	setp.ne.s32 	%p80, %r9, %r118;
	@%p80 bra 	$Lt_22_168450;
	.loc	20	59	0
	mov.f32 	%f112, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_29;
$Lt_22_168450:
	.loc	20	61	0
	mov.u32 	%r119, 64;
	setp.ne.s32 	%p81, %r9, %r119;
	@%p81 bra 	$Lt_22_168706;
	.loc	20	63	0
	mov.f32 	%f112, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_29;
$Lt_22_168706:
	.loc	20	65	0
	mov.u32 	%r120, 128;
	setp.ne.s32 	%p82, %r9, %r120;
	@%p82 bra 	$Lt_22_168962;
	.loc	20	68	0
	mov.f32 	%f112, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_29;
$Lt_22_168962:
	.loc	20	70	0
	mov.u32 	%r121, 192;
	setp.ne.s32 	%p83, %r9, %r121;
	@%p83 bra 	$Lt_22_169218;
	.loc	20	72	0
	mov.f32 	%f112, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_29;
$Lt_22_169218:
	.loc	20	76	0
	mov.f32 	%f112, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_199_29:
	.loc	20	107	0
	and.b32 	%r122, %r4, 2048;
	mov.s32 	%r123, 0;
	setp.ne.s32 	%p84, %r122, %r123;
	@!%p84 bra 	$Lt_22_237314;
	.loc	20	100	0
	ld.const.f32 	%f113, [kYCbCrOffset+0];
	bra.uni 	$Lt_22_237058;
$Lt_22_237314:
	ld.const.f32 	%f113, [kYCbCrFullRangeOffset+0];
$Lt_22_237058:
	.loc	20	107	0
	@!%p84 bra 	$Lt_22_237826;
	.loc	20	100	0
	ld.const.f32 	%f114, [kYCbCrOffset+4];
	bra.uni 	$Lt_22_237570;
$Lt_22_237826:
	ld.const.f32 	%f114, [kYCbCrFullRangeOffset+4];
$Lt_22_237570:
	.loc	20	107	0
	@!%p84 bra 	$Lt_22_238338;
	.loc	20	100	0
	ld.const.f32 	%f115, [kYCbCrOffset+8];
	bra.uni 	$Lt_22_238082;
$Lt_22_238338:
	ld.const.f32 	%f115, [kYCbCrFullRangeOffset+8];
$Lt_22_238082:
	.loc	20	527	0
	mov.f32 	%f116, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f117, %f112, %f116;
	fma.rn.ftz.f32 	%f10, %f117, %f113, %f10;
	fma.rn.ftz.f32 	%f11, %f117, %f114, %f11;
	fma.rn.ftz.f32 	%f12, %f117, %f115, %f12;
$Lt_22_253698:
$Lt_22_91650:
	.loc	20	525	0
	and.b32 	%r124, %r2, 12;
	and.b32 	%r125, %r4, 12;
	setp.eq.s32 	%p85, %r124, %r125;
	@%p85 bra 	$Lt_22_239106;
	.loc	20	532	0
	mov.u32 	%r126, 8;
	setp.ne.s32 	%p86, %r124, %r126;
	@%p86 bra 	$L_22_222466;
	mov.u32 	%r127, 12;
	setp.eq.s32 	%p87, %r125, %r127;
	@%p87 bra 	$Lt_22_254466;
$L_22_222466:
	mov.u32 	%r128, 12;
	setp.eq.s32 	%p88, %r124, %r128;
	@%p88 bra 	$Lt_22_254466;
	mov.u32 	%r129, 0;
	setp.ne.s32 	%p89, %r124, %r129;
	@%p89 bra 	$L_22_221698;
$Lt_22_254466:
$L_22_221954:
	.loc	20	57	0
	mov.u32 	%r130, 0;
	setp.ne.s32 	%p90, %r9, %r130;
	@%p90 bra 	$Lt_22_171778;
	.loc	20	59	0
	mov.f32 	%f118, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_25;
$Lt_22_171778:
	.loc	20	61	0
	mov.u32 	%r131, 64;
	setp.ne.s32 	%p91, %r9, %r131;
	@%p91 bra 	$Lt_22_172034;
	.loc	20	63	0
	mov.f32 	%f118, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_25;
$Lt_22_172034:
	.loc	20	65	0
	mov.u32 	%r132, 128;
	setp.ne.s32 	%p92, %r9, %r132;
	@%p92 bra 	$Lt_22_172290;
	.loc	20	68	0
	mov.f32 	%f118, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_25;
$Lt_22_172290:
	.loc	20	70	0
	mov.u32 	%r133, 192;
	setp.ne.s32 	%p93, %r9, %r133;
	@%p93 bra 	$Lt_22_172546;
	.loc	20	72	0
	mov.f32 	%f118, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_25;
$Lt_22_172546:
	.loc	20	76	0
	mov.f32 	%f118, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_199_25:
	.loc	20	536	0
	mov.f32 	%f9, %f118;
	bra.uni 	$Lt_22_239106;
$L_22_221698:
	.loc	20	540	0
	mov.s32 	%r134, 12;
	setp.eq.s32 	%p94, %r125, %r134;
	mov.s32 	%r135, 4;
	set.eq.u32.s32 	%r136, %r124, %r135;
	neg.s32 	%r137, %r136;
	selp.s32 	%r138, 1, 0, %p94;
	mov.s32 	%r139, 8;
	set.eq.u32.s32 	%r140, %r125, %r139;
	neg.s32 	%r141, %r140;
	or.b32 	%r142, %r138, %r141;
	and.b32 	%r143, %r137, %r142;
	mov.u32 	%r144, 0;
	setp.eq.s32 	%p95, %r143, %r144;
	@%p95 bra 	$Lt_22_239362;
	.loc	20	410	0
	mov.f32 	%f119, %f10;
	mov.f32 	%f120, %f119;
	mov.f32 	%f121, %f11;
	mov.f32 	%f122, %f121;
	mov.f32 	%f123, %f12;
	mov.f32 	%f124, %f123;
	.loc	20	57	0
	mov.s32 	%r145, 0;
	setp.eq.s32 	%p35, %r9, %r145;
	@!%p35 bra 	$Lt_22_173058;
	.loc	20	59	0
	mov.f32 	%f125, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_23;
$Lt_22_173058:
	.loc	20	61	0
	mov.u32 	%r146, 64;
	setp.ne.s32 	%p96, %r9, %r146;
	@%p96 bra 	$Lt_22_173314;
	.loc	20	63	0
	mov.f32 	%f125, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_23;
$Lt_22_173314:
	.loc	20	65	0
	mov.u32 	%r147, 128;
	setp.ne.s32 	%p97, %r9, %r147;
	@%p97 bra 	$Lt_22_173570;
	.loc	20	68	0
	mov.f32 	%f125, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_23;
$Lt_22_173570:
	.loc	20	70	0
	mov.u32 	%r148, 192;
	setp.ne.s32 	%p98, %r9, %r148;
	@%p98 bra 	$Lt_22_173826;
	.loc	20	72	0
	mov.f32 	%f125, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_23;
$Lt_22_173826:
	.loc	20	76	0
	mov.f32 	%f125, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_199_23:
	.loc	20	413	0
	mov.u32 	%r149, 0;
	setp.eq.s32 	%p99, %r18, %r149;
	@%p99 bra 	$Lt_22_239618;
	.loc	20	57	0
	@!%p35 bra 	$Lt_22_174338;
	.loc	20	59	0
	mov.f32 	%f126, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_21;
$Lt_22_174338:
	.loc	20	61	0
	mov.u32 	%r150, 64;
	setp.ne.s32 	%p100, %r9, %r150;
	@%p100 bra 	$Lt_22_174594;
	.loc	20	63	0
	mov.f32 	%f126, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_21;
$Lt_22_174594:
	.loc	20	65	0
	mov.u32 	%r151, 128;
	setp.ne.s32 	%p101, %r9, %r151;
	@%p101 bra 	$Lt_22_174850;
	.loc	20	68	0
	mov.f32 	%f126, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_21;
$Lt_22_174850:
	.loc	20	70	0
	mov.u32 	%r152, 192;
	setp.ne.s32 	%p102, %r9, %r152;
	@%p102 bra 	$Lt_22_175106;
	.loc	20	72	0
	mov.f32 	%f126, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_21;
$Lt_22_175106:
	.loc	20	76	0
	mov.f32 	%f126, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_199_21:
	.loc	20	118	0
	and.b32 	%r122, %r4, 2048;
	mov.s32 	%r153, 0;
	setp.ne.s32 	%p84, %r122, %r153;
	@!%p84 bra 	$Lt_22_240386;
	.loc	20	100	0
	ld.const.f32 	%f127, [kYCbCrOffset+0];
	bra.uni 	$Lt_22_240130;
$Lt_22_240386:
	ld.const.f32 	%f127, [kYCbCrFullRangeOffset+0];
$Lt_22_240130:
	.loc	20	118	0
	@!%p84 bra 	$Lt_22_240898;
	.loc	20	100	0
	ld.const.f32 	%f128, [kYCbCrOffset+4];
	bra.uni 	$Lt_22_240642;
$Lt_22_240898:
	ld.const.f32 	%f128, [kYCbCrFullRangeOffset+4];
$Lt_22_240642:
	.loc	20	118	0
	@!%p84 bra 	$Lt_22_241410;
	.loc	20	100	0
	ld.const.f32 	%f129, [kYCbCrOffset+8];
	bra.uni 	$Lt_22_241154;
$Lt_22_241410:
	ld.const.f32 	%f129, [kYCbCrFullRangeOffset+8];
$Lt_22_241154:
	.loc	20	415	0
	mov.f32 	%f130, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f131, %f126, %f130;
	mul.ftz.f32 	%f132, %f131, %f127;
	sub.ftz.f32 	%f120, %f119, %f132;
	mul.ftz.f32 	%f133, %f131, %f128;
	sub.ftz.f32 	%f122, %f121, %f133;
	mul.ftz.f32 	%f134, %f131, %f129;
	sub.ftz.f32 	%f124, %f123, %f134;
$Lt_22_239618:
	.loc	20	418	0
	rcp.approx.ftz.f32 	%f135, %f125;
	mul.ftz.f32 	%f136, %f135, %f9;
	mul.ftz.f32 	%f120, %f136, %f120;
	.loc	20	419	0
	mul.ftz.f32 	%f122, %f136, %f122;
	.loc	20	420	0
	mul.ftz.f32 	%f124, %f136, %f124;
	.loc	20	422	0
	mov.u32 	%r154, 0;
	setp.eq.s32 	%p103, %r18, %r154;
	@%p103 bra 	$Lt_22_241666;
	.loc	20	57	0
	@!%p35 bra 	$Lt_22_177410;
	.loc	20	59	0
	mov.f32 	%f137, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_17;
$Lt_22_177410:
	.loc	20	61	0
	mov.u32 	%r155, 64;
	setp.ne.s32 	%p104, %r9, %r155;
	@%p104 bra 	$Lt_22_177666;
	.loc	20	63	0
	mov.f32 	%f137, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_17;
$Lt_22_177666:
	.loc	20	65	0
	mov.u32 	%r156, 128;
	setp.ne.s32 	%p105, %r9, %r156;
	@%p105 bra 	$Lt_22_177922;
	.loc	20	68	0
	mov.f32 	%f137, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_17;
$Lt_22_177922:
	.loc	20	70	0
	mov.u32 	%r157, 192;
	setp.ne.s32 	%p106, %r9, %r157;
	@%p106 bra 	$Lt_22_178178;
	.loc	20	72	0
	mov.f32 	%f137, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_17;
$Lt_22_178178:
	.loc	20	76	0
	mov.f32 	%f137, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_199_17:
	.loc	20	107	0
	and.b32 	%r122, %r4, 2048;
	mov.s32 	%r158, 0;
	setp.ne.s32 	%p84, %r122, %r158;
	@!%p84 bra 	$Lt_22_242434;
	.loc	20	100	0
	ld.const.f32 	%f138, [kYCbCrOffset+0];
	bra.uni 	$Lt_22_242178;
$Lt_22_242434:
	ld.const.f32 	%f138, [kYCbCrFullRangeOffset+0];
$Lt_22_242178:
	.loc	20	107	0
	@!%p84 bra 	$Lt_22_242946;
	.loc	20	100	0
	ld.const.f32 	%f139, [kYCbCrOffset+4];
	bra.uni 	$Lt_22_242690;
$Lt_22_242946:
	ld.const.f32 	%f139, [kYCbCrFullRangeOffset+4];
$Lt_22_242690:
	.loc	20	107	0
	@!%p84 bra 	$Lt_22_243458;
	.loc	20	100	0
	ld.const.f32 	%f140, [kYCbCrOffset+8];
	bra.uni 	$Lt_22_243202;
$Lt_22_243458:
	ld.const.f32 	%f140, [kYCbCrFullRangeOffset+8];
$Lt_22_243202:
	.loc	20	424	0
	mov.f32 	%f141, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f142, %f137, %f141;
	fma.rn.ftz.f32 	%f120, %f142, %f138, %f120;
	fma.rn.ftz.f32 	%f122, %f142, %f139, %f122;
	fma.rn.ftz.f32 	%f124, %f142, %f140, %f124;
$Lt_22_241666:
	.loc	20	543	0
	mov.f32 	%f10, %f120;
	mov.f32 	%f11, %f122;
	mov.f32 	%f12, %f124;
	@!%p94 bra 	$Lt_22_239106;
	.loc	20	57	0
	@!%p35 bra 	$Lt_22_180482;
	.loc	20	59	0
	mov.f32 	%f118, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_13;
$Lt_22_180482:
	.loc	20	61	0
	mov.u32 	%r159, 64;
	setp.ne.s32 	%p107, %r9, %r159;
	@%p107 bra 	$Lt_22_180738;
	.loc	20	63	0
	mov.f32 	%f118, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_13;
$Lt_22_180738:
	.loc	20	65	0
	mov.u32 	%r160, 128;
	setp.ne.s32 	%p108, %r9, %r160;
	@%p108 bra 	$Lt_22_180994;
	.loc	20	68	0
	mov.f32 	%f118, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_13;
$Lt_22_180994:
	.loc	20	70	0
	mov.u32 	%r161, 192;
	setp.ne.s32 	%p109, %r9, %r161;
	@%p109 bra 	$Lt_22_181250;
	.loc	20	72	0
	mov.f32 	%f118, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_13;
$Lt_22_181250:
	.loc	20	76	0
	mov.f32 	%f118, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_199_13:
	.loc	20	546	0
	mov.f32 	%f9, %f118;
	bra.uni 	$Lt_22_239106;
$Lt_22_239362:
	.loc	20	433	0
	mov.f32 	%f143, %f9;
	mov.f32 	%f144, %f10;
	mov.f32 	%f145, %f144;
	mov.f32 	%f146, %f11;
	mov.f32 	%f147, %f146;
	mov.f32 	%f148, %f12;
	mov.f32 	%f149, %f148;
	.loc	20	435	0
	mov.u32 	%r162, 0;
	setp.eq.s32 	%p110, %r18, %r162;
	@%p110 bra 	$Lt_22_244226;
	.loc	20	57	0
	mov.u32 	%r163, 0;
	setp.ne.s32 	%p111, %r9, %r163;
	@%p111 bra 	$Lt_22_181762;
	.loc	20	59	0
	mov.f32 	%f150, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_11;
$Lt_22_181762:
	.loc	20	61	0
	mov.u32 	%r164, 64;
	setp.ne.s32 	%p112, %r9, %r164;
	@%p112 bra 	$Lt_22_182018;
	.loc	20	63	0
	mov.f32 	%f150, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_11;
$Lt_22_182018:
	.loc	20	65	0
	mov.u32 	%r165, 128;
	setp.ne.s32 	%p113, %r9, %r165;
	@%p113 bra 	$Lt_22_182274;
	.loc	20	68	0
	mov.f32 	%f150, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_11;
$Lt_22_182274:
	.loc	20	70	0
	mov.u32 	%r166, 192;
	setp.ne.s32 	%p114, %r9, %r166;
	@%p114 bra 	$Lt_22_182530;
	.loc	20	72	0
	mov.f32 	%f150, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_11;
$Lt_22_182530:
	.loc	20	76	0
	mov.f32 	%f150, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_199_11:
	.loc	20	118	0
	and.b32 	%r122, %r4, 2048;
	mov.s32 	%r167, 0;
	setp.ne.s32 	%p84, %r122, %r167;
	@!%p84 bra 	$Lt_22_244994;
	.loc	20	100	0
	ld.const.f32 	%f151, [kYCbCrOffset+0];
	bra.uni 	$Lt_22_244738;
$Lt_22_244994:
	ld.const.f32 	%f151, [kYCbCrFullRangeOffset+0];
$Lt_22_244738:
	.loc	20	118	0
	@!%p84 bra 	$Lt_22_245506;
	.loc	20	100	0
	ld.const.f32 	%f152, [kYCbCrOffset+4];
	bra.uni 	$Lt_22_245250;
$Lt_22_245506:
	ld.const.f32 	%f152, [kYCbCrFullRangeOffset+4];
$Lt_22_245250:
	.loc	20	118	0
	@!%p84 bra 	$Lt_22_246018;
	.loc	20	100	0
	ld.const.f32 	%f153, [kYCbCrOffset+8];
	bra.uni 	$Lt_22_245762;
$Lt_22_246018:
	ld.const.f32 	%f153, [kYCbCrFullRangeOffset+8];
$Lt_22_245762:
	.loc	20	437	0
	mov.f32 	%f154, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f155, %f150, %f154;
	mul.ftz.f32 	%f156, %f155, %f151;
	sub.ftz.f32 	%f145, %f144, %f156;
	mul.ftz.f32 	%f157, %f155, %f152;
	sub.ftz.f32 	%f147, %f146, %f157;
	mul.ftz.f32 	%f158, %f155, %f153;
	sub.ftz.f32 	%f149, %f148, %f158;
$Lt_22_244226:
	mov.f32 	%f159, 0fb70637bd;   	// -8e-006
	add.ftz.f32 	%f160, %f9, %f159;
	mov.f32 	%f161, 0f00000000;   	// 0
	setp.le.ftz.f32 	%p115, %f160, %f161;
	@!%p115 bra 	$Lt_22_246530;
	mov.f32 	%f149, 0f00000000;   	// 0
	mov.f32 	%f147, 0f00000000;   	// 0
	mov.f32 	%f145, 0f00000000;   	// 0
	mov.f32 	%f143, 0f00000000;   	// 0
	bra.uni 	$Lt_22_246274;
$Lt_22_246530:
	.loc	20	57	0
	mov.u32 	%r168, 0;
	setp.ne.s32 	%p116, %r9, %r168;
	@%p116 bra 	$Lt_22_184834;
	.loc	20	59	0
	mov.f32 	%f162, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_7;
$Lt_22_184834:
	.loc	20	61	0
	mov.u32 	%r169, 64;
	setp.ne.s32 	%p117, %r9, %r169;
	@%p117 bra 	$Lt_22_185090;
	.loc	20	63	0
	mov.f32 	%f162, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_7;
$Lt_22_185090:
	.loc	20	65	0
	mov.u32 	%r170, 128;
	setp.ne.s32 	%p118, %r9, %r170;
	@%p118 bra 	$Lt_22_185346;
	.loc	20	68	0
	mov.f32 	%f162, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_7;
$Lt_22_185346:
	.loc	20	70	0
	mov.u32 	%r171, 192;
	setp.ne.s32 	%p119, %r9, %r171;
	@%p119 bra 	$Lt_22_185602;
	.loc	20	72	0
	mov.f32 	%f162, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_7;
$Lt_22_185602:
	.loc	20	76	0
	mov.f32 	%f162, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_199_7:
	.loc	20	447	0
	div.approx.ftz.f32 	%f163, %f162, %f9;
	mul.ftz.f32 	%f145, %f163, %f145;
	.loc	20	448	0
	mul.ftz.f32 	%f147, %f163, %f147;
	.loc	20	449	0
	mul.ftz.f32 	%f149, %f163, %f149;
$Lt_22_246274:
	.loc	20	452	0
	mov.u32 	%r172, 0;
	setp.eq.s32 	%p120, %r18, %r172;
	@%p120 bra 	$Lt_22_246786;
	.loc	20	57	0
	mov.u32 	%r173, 0;
	setp.ne.s32 	%p121, %r9, %r173;
	@%p121 bra 	$Lt_22_186114;
	.loc	20	59	0
	mov.f32 	%f164, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_5;
$Lt_22_186114:
	.loc	20	61	0
	mov.u32 	%r174, 64;
	setp.ne.s32 	%p122, %r9, %r174;
	@%p122 bra 	$Lt_22_186370;
	.loc	20	63	0
	mov.f32 	%f164, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_5;
$Lt_22_186370:
	.loc	20	65	0
	mov.u32 	%r175, 128;
	setp.ne.s32 	%p123, %r9, %r175;
	@%p123 bra 	$Lt_22_186626;
	.loc	20	68	0
	mov.f32 	%f164, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_5;
$Lt_22_186626:
	.loc	20	70	0
	mov.u32 	%r176, 192;
	setp.ne.s32 	%p124, %r9, %r176;
	@%p124 bra 	$Lt_22_186882;
	.loc	20	72	0
	mov.f32 	%f164, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_5;
$Lt_22_186882:
	.loc	20	76	0
	mov.f32 	%f164, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_199_5:
	.loc	20	107	0
	and.b32 	%r122, %r4, 2048;
	mov.s32 	%r177, 0;
	setp.ne.s32 	%p84, %r122, %r177;
	@!%p84 bra 	$Lt_22_247554;
	.loc	20	100	0
	ld.const.f32 	%f165, [kYCbCrOffset+0];
	bra.uni 	$Lt_22_247298;
$Lt_22_247554:
	ld.const.f32 	%f165, [kYCbCrFullRangeOffset+0];
$Lt_22_247298:
	.loc	20	107	0
	@!%p84 bra 	$Lt_22_248066;
	.loc	20	100	0
	ld.const.f32 	%f166, [kYCbCrOffset+4];
	bra.uni 	$Lt_22_247810;
$Lt_22_248066:
	ld.const.f32 	%f166, [kYCbCrFullRangeOffset+4];
$Lt_22_247810:
	.loc	20	107	0
	@!%p84 bra 	$Lt_22_248578;
	.loc	20	100	0
	ld.const.f32 	%f167, [kYCbCrOffset+8];
	bra.uni 	$Lt_22_248322;
$Lt_22_248578:
	ld.const.f32 	%f167, [kYCbCrFullRangeOffset+8];
$Lt_22_248322:
	.loc	20	454	0
	mov.f32 	%f168, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f169, %f164, %f168;
	fma.rn.ftz.f32 	%f145, %f169, %f165, %f145;
	fma.rn.ftz.f32 	%f147, %f169, %f166, %f147;
	fma.rn.ftz.f32 	%f149, %f169, %f167, %f149;
$Lt_22_246786:
	.loc	20	551	0
	mov.f32 	%f9, %f143;
	mov.f32 	%f10, %f145;
	mov.f32 	%f11, %f147;
	mov.f32 	%f12, %f149;
$Lt_22_239106:
$L_22_221442:
$Lt_22_238594:
	.loc	20	540	0
	and.b32 	%r178, %r4, 4096;
	mov.u32 	%r179, 0;
	setp.ne.s32 	%p125, %r178, %r179;
	@%p125 bra 	$Lt_22_248834;
	.loc	21	268	0
	mov.f32 	%f170, %f10;
	.loc	21	269	0
	mov.f32 	%f171, %f9;
	.loc	20	558	0
	mov.f32 	%f9, %f12;
	mov.f32 	%f10, %f11;
	mov.f32 	%f11, %f170;
	mov.f32 	%f12, %f171;
$Lt_22_248834:
	@!%p3 bra 	$Lt_22_249346;
	.loc	20	57	0
	mov.u32 	%r180, 0;
	setp.ne.s32 	%p126, %r9, %r180;
	@%p126 bra 	$Lt_22_189442;
	.loc	20	59	0
	mov.f32 	%f172, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_1;
$Lt_22_189442:
	.loc	20	61	0
	mov.u32 	%r181, 64;
	setp.ne.s32 	%p127, %r9, %r181;
	@%p127 bra 	$Lt_22_189698;
	.loc	20	63	0
	mov.f32 	%f172, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_1;
$Lt_22_189698:
	.loc	20	65	0
	mov.u32 	%r182, 128;
	setp.ne.s32 	%p128, %r9, %r182;
	@%p128 bra 	$Lt_22_189954;
	.loc	20	68	0
	mov.f32 	%f172, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_1;
$Lt_22_189954:
	.loc	20	70	0
	mov.u32 	%r183, 192;
	setp.ne.s32 	%p129, %r9, %r183;
	@%p129 bra 	$Lt_22_190210;
	.loc	20	72	0
	mov.f32 	%f172, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_199_1;
$Lt_22_190210:
	.loc	20	76	0
	mov.f32 	%f172, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_199_1:
	.loc	20	564	0
	mov.f32 	%f173, 0f3f000000;   	// 0.5
	add.ftz.f32 	%f174, %f9, %f173;
	mov.f32 	%f175, 0f00000000;   	// 0
	mov.f32 	%f176, 0f00000000;   	// 0
	setp.gt.ftz.f32 	%p130, %f174, %f176;
	selp.f32 	%f177, %f174, %f175, %p130;
	min.ftz.f32 	%f9, %f177, %f172;
	mov.f32 	%f178, 0f3f000000;   	// 0.5
	add.ftz.f32 	%f179, %f10, %f178;
	mov.f32 	%f180, 0f00000000;   	// 0
	mov.f32 	%f181, 0f00000000;   	// 0
	setp.gt.ftz.f32 	%p131, %f179, %f181;
	selp.f32 	%f182, %f179, %f180, %p131;
	min.ftz.f32 	%f10, %f182, %f172;
	mov.f32 	%f183, 0f3f000000;   	// 0.5
	add.ftz.f32 	%f184, %f11, %f183;
	mov.f32 	%f185, 0f00000000;   	// 0
	mov.f32 	%f186, 0f00000000;   	// 0
	setp.gt.ftz.f32 	%p132, %f184, %f186;
	selp.f32 	%f187, %f184, %f185, %p132;
	min.ftz.f32 	%f11, %f187, %f172;
	mov.f32 	%f188, 0f3f000000;   	// 0.5
	add.ftz.f32 	%f189, %f12, %f188;
	mov.f32 	%f190, 0f00000000;   	// 0
	mov.f32 	%f191, 0f00000000;   	// 0
	setp.gt.ftz.f32 	%p133, %f189, %f191;
	selp.f32 	%f192, %f189, %f190, %p133;
	min.ftz.f32 	%f12, %f192, %f172;
$Lt_22_249346:
	.loc	20	567	0
	mov.f32 	%f193, %f9;
	st.param.f32 	[__cudaretf__Z23ConvertPixel_444_To_4446float414IR_PixelFormatS0_+0], %f193;
	mov.f32 	%f194, %f10;
	st.param.f32 	[__cudaretf__Z23ConvertPixel_444_To_4446float414IR_PixelFormatS0_+4], %f194;
	mov.f32 	%f195, %f11;
	st.param.f32 	[__cudaretf__Z23ConvertPixel_444_To_4446float414IR_PixelFormatS0_+8], %f195;
	mov.f32 	%f196, %f12;
	st.param.f32 	[__cudaretf__Z23ConvertPixel_444_To_4446float414IR_PixelFormatS0_+12], %f196;
	ret;
$LDWend__Z23ConvertPixel_444_To_4446float414IR_PixelFormatS0_:
	} // _Z23ConvertPixel_444_To_4446float414IR_PixelFormatS0_

	.visible .func (.param .align 16 .b8 __cudaretf__Z18SwapComponentOrderI6float4ET_RKS1_[16]) _Z18SwapComponentOrderI6float4ET_RKS1_ (.param .u64 __cudaparmf1__Z18SwapComponentOrderI6float4ET_RKS1_)
	{
	.reg .u64 %rd<4>;
	.reg .f32 %f<6>;
	.loc	21	264	0
$LDWbegin__Z18SwapComponentOrderI6float4ET_RKS1_:
	ld.param.u64 	%rd1, [__cudaparmf1__Z18SwapComponentOrderI6float4ET_RKS1_];
	mov.s64 	%rd2, %rd1;
	.loc	21	270	0
	ld.f32 	%f1, [%rd2+12];
	st.param.f32 	[__cudaretf__Z18SwapComponentOrderI6float4ET_RKS1_+0], %f1;
	ld.f32 	%f2, [%rd2+8];
	st.param.f32 	[__cudaretf__Z18SwapComponentOrderI6float4ET_RKS1_+4], %f2;
	ld.f32 	%f3, [%rd2+4];
	st.param.f32 	[__cudaretf__Z18SwapComponentOrderI6float4ET_RKS1_+8], %f3;
	ld.f32 	%f4, [%rd2+0];
	st.param.f32 	[__cudaretf__Z18SwapComponentOrderI6float4ET_RKS1_+12], %f4;
	ret;
$LDWend__Z18SwapComponentOrderI6float4ET_RKS1_:
	} // _Z18SwapComponentOrderI6float4ET_RKS1_

	.visible .func (.param .align 16 .b8 __cudaretf__ZplI6float4ET_RKS1_f[16]) _ZplI6float4ET_RKS1_f (.param .u64 __cudaparmf1__ZplI6float4ET_RKS1_f, .param .f32 __cudaparmf2__ZplI6float4ET_RKS1_f)
	{
	.reg .u64 %rd<4>;
	.reg .f32 %f<12>;
	.loc	21	113	0
$LDWbegin__ZplI6float4ET_RKS1_f:
	ld.param.u64 	%rd1, [__cudaparmf1__ZplI6float4ET_RKS1_f];
	mov.s64 	%rd2, %rd1;
	ld.param.f32 	%f1, [__cudaparmf2__ZplI6float4ET_RKS1_f];
	mov.f32 	%f2, %f1;
	ld.v4.f32 	{%f3,%f4,%f5,%f6}, [%rd2+0];
	.loc	21	116	0
	add.ftz.f32 	%f7, %f3, %f2;
	st.param.f32 	[__cudaretf__ZplI6float4ET_RKS1_f+0], %f7;
	add.ftz.f32 	%f8, %f4, %f2;
	st.param.f32 	[__cudaretf__ZplI6float4ET_RKS1_f+4], %f8;
	add.ftz.f32 	%f9, %f5, %f2;
	st.param.f32 	[__cudaretf__ZplI6float4ET_RKS1_f+8], %f9;
	add.ftz.f32 	%f10, %f6, %f2;
	st.param.f32 	[__cudaretf__ZplI6float4ET_RKS1_f+12], %f10;
	ret;
$LDWend__ZplI6float4ET_RKS1_f:
	} // _ZplI6float4ET_RKS1_f

	.visible .func (.param .align 16 .b8 __cudaretf__Z15ClampComponentsI6float4ET_RKS1_ff[16]) _Z15ClampComponentsI6float4ET_RKS1_ff (.param .u64 __cudaparmf1__Z15ClampComponentsI6float4ET_RKS1_ff, .param .f32 __cudaparmf2__Z15ClampComponentsI6float4ET_RKS1_ff, .param .f32 __cudaparmf3__Z15ClampComponentsI6float4ET_RKS1_ff)
	{
	.reg .u64 %rd<4>;
	.reg .f32 %f<18>;
	.reg .pred %p<6>;
	.loc	21	342	0
$LDWbegin__Z15ClampComponentsI6float4ET_RKS1_ff:
	ld.param.u64 	%rd1, [__cudaparmf1__Z15ClampComponentsI6float4ET_RKS1_ff];
	mov.s64 	%rd2, %rd1;
	ld.param.f32 	%f1, [__cudaparmf2__Z15ClampComponentsI6float4ET_RKS1_ff];
	mov.f32 	%f2, %f1;
	ld.param.f32 	%f3, [__cudaparmf3__Z15ClampComponentsI6float4ET_RKS1_ff];
	mov.f32 	%f4, %f3;
	.loc	21	344	0
	ld.f32 	%f5, [%rd2+0];
	.loc	21	345	0
	ld.f32 	%f6, [%rd2+4];
	.loc	21	346	0
	ld.f32 	%f7, [%rd2+8];
	.loc	21	347	0
	ld.f32 	%f8, [%rd2+12];
	.loc	21	348	0
	setp.lt.ftz.f32 	%p1, %f2, %f5;
	selp.f32 	%f9, %f5, %f2, %p1;
	min.ftz.f32 	%f10, %f9, %f4;
	st.param.f32 	[__cudaretf__Z15ClampComponentsI6float4ET_RKS1_ff+0], %f10;
	setp.lt.ftz.f32 	%p2, %f2, %f6;
	selp.f32 	%f11, %f6, %f2, %p2;
	min.ftz.f32 	%f12, %f11, %f4;
	st.param.f32 	[__cudaretf__Z15ClampComponentsI6float4ET_RKS1_ff+4], %f12;
	setp.lt.ftz.f32 	%p3, %f2, %f7;
	selp.f32 	%f13, %f7, %f2, %p3;
	min.ftz.f32 	%f14, %f13, %f4;
	st.param.f32 	[__cudaretf__Z15ClampComponentsI6float4ET_RKS1_ff+8], %f14;
	setp.lt.ftz.f32 	%p4, %f2, %f8;
	selp.f32 	%f15, %f8, %f2, %p4;
	min.ftz.f32 	%f16, %f15, %f4;
	st.param.f32 	[__cudaretf__Z15ClampComponentsI6float4ET_RKS1_ff+12], %f16;
	ret;
$LDWend__Z15ClampComponentsI6float4ET_RKS1_ff:
	} // _Z15ClampComponentsI6float4ET_RKS1_ff

	.visible .func _Z7Write2DI7ushort4EvT_PS1_iii (.param .align 8 .b8 __cudaparmf1__Z7Write2DI7ushort4EvT_PS1_iii[8], .param .u64 __cudaparmf2__Z7Write2DI7ushort4EvT_PS1_iii, .param .s32 __cudaparmf3__Z7Write2DI7ushort4EvT_PS1_iii, .param .s32 __cudaparmf4__Z7Write2DI7ushort4EvT_PS1_iii, .param .s32 __cudaparmf5__Z7Write2DI7ushort4EvT_PS1_iii)
	{
	.reg .u32 %r<18>;
	.reg .u64 %rd<7>;
	.loc	19	125	0
$LDWbegin__Z7Write2DI7ushort4EvT_PS1_iii:
	ld.param.u16 	%r1, [__cudaparmf1__Z7Write2DI7ushort4EvT_PS1_iii+0];
	mov.s32 	%r2, %r1;
	ld.param.u16 	%r3, [__cudaparmf1__Z7Write2DI7ushort4EvT_PS1_iii+2];
	mov.s32 	%r4, %r3;
	ld.param.u16 	%r5, [__cudaparmf1__Z7Write2DI7ushort4EvT_PS1_iii+4];
	mov.s32 	%r6, %r5;
	ld.param.u16 	%r7, [__cudaparmf1__Z7Write2DI7ushort4EvT_PS1_iii+6];
	mov.s32 	%r8, %r7;
	ld.param.u64 	%rd1, [__cudaparmf2__Z7Write2DI7ushort4EvT_PS1_iii];
	mov.s64 	%rd2, %rd1;
	ld.param.u32 	%r9, [__cudaparmf3__Z7Write2DI7ushort4EvT_PS1_iii];
	mov.s32 	%r10, %r9;
	ld.param.u32 	%r11, [__cudaparmf4__Z7Write2DI7ushort4EvT_PS1_iii];
	mov.s32 	%r12, %r11;
	ld.param.u32 	%r13, [__cudaparmf5__Z7Write2DI7ushort4EvT_PS1_iii];
	mov.s32 	%r14, %r13;
	.loc	19	126	0
	mul.lo.s32 	%r15, %r10, %r14;
	add.s32 	%r16, %r12, %r15;
	cvt.s64.s32 	%rd3, %r16;
	mul.wide.s32 	%rd4, %r16, 8;
	add.u64 	%rd5, %rd2, %rd4;
	st.v4.u16 	[%rd5+0], {%r2,%r4,%r6,%r8};
	.loc	19	127	0
	ret;
$LDWend__Z7Write2DI7ushort4EvT_PS1_iii:
	} // _Z7Write2DI7ushort4EvT_PS1_iii

	.visible .func _Z7Write2DI6float4EvT_PS1_iii (.param .align 16 .b8 __cudaparmf1__Z7Write2DI6float4EvT_PS1_iii[16], .param .u64 __cudaparmf2__Z7Write2DI6float4EvT_PS1_iii, .param .s32 __cudaparmf3__Z7Write2DI6float4EvT_PS1_iii, .param .s32 __cudaparmf4__Z7Write2DI6float4EvT_PS1_iii, .param .s32 __cudaparmf5__Z7Write2DI6float4EvT_PS1_iii)
	{
	.reg .u32 %r<10>;
	.reg .u64 %rd<7>;
	.reg .f32 %f<10>;
	.loc	19	125	0
$LDWbegin__Z7Write2DI6float4EvT_PS1_iii:
	ld.param.f32 	%f1, [__cudaparmf1__Z7Write2DI6float4EvT_PS1_iii+0];
	mov.f32 	%f2, %f1;
	ld.param.f32 	%f3, [__cudaparmf1__Z7Write2DI6float4EvT_PS1_iii+4];
	mov.f32 	%f4, %f3;
	ld.param.f32 	%f5, [__cudaparmf1__Z7Write2DI6float4EvT_PS1_iii+8];
	mov.f32 	%f6, %f5;
	ld.param.f32 	%f7, [__cudaparmf1__Z7Write2DI6float4EvT_PS1_iii+12];
	mov.f32 	%f8, %f7;
	ld.param.u64 	%rd1, [__cudaparmf2__Z7Write2DI6float4EvT_PS1_iii];
	mov.s64 	%rd2, %rd1;
	ld.param.u32 	%r1, [__cudaparmf3__Z7Write2DI6float4EvT_PS1_iii];
	mov.s32 	%r2, %r1;
	ld.param.u32 	%r3, [__cudaparmf4__Z7Write2DI6float4EvT_PS1_iii];
	mov.s32 	%r4, %r3;
	ld.param.u32 	%r5, [__cudaparmf5__Z7Write2DI6float4EvT_PS1_iii];
	mov.s32 	%r6, %r5;
	.loc	19	126	0
	mul.lo.s32 	%r7, %r2, %r6;
	add.s32 	%r8, %r4, %r7;
	cvt.s64.s32 	%rd3, %r8;
	mul.wide.s32 	%rd4, %r8, 16;
	add.u64 	%rd5, %rd2, %rd4;
	st.v4.f32 	[%rd5+0], {%f2,%f4,%f6,%f8};
	.loc	19	127	0
	ret;
$LDWend__Z7Write2DI6float4EvT_PS1_iii:
	} // _Z7Write2DI6float4EvT_PS1_iii

	.visible .func (.param .align 16 .b8 __cudaretf__Z18UnpremultiplyPixel8PixelRGB[16]) _Z18UnpremultiplyPixel8PixelRGB (.param .align 16 .b8 __cudaparmf1__Z18UnpremultiplyPixel8PixelRGB[16])
	{
	.reg .f32 %f<23>;
	.reg .pred %p<3>;
	.loc	4	206	0
$LDWbegin__Z18UnpremultiplyPixel8PixelRGB:
	ld.param.f32 	%f1, [__cudaparmf1__Z18UnpremultiplyPixel8PixelRGB+0];
	mov.f32 	%f2, %f1;
	ld.param.f32 	%f3, [__cudaparmf1__Z18UnpremultiplyPixel8PixelRGB+4];
	mov.f32 	%f4, %f3;
	ld.param.f32 	%f5, [__cudaparmf1__Z18UnpremultiplyPixel8PixelRGB+8];
	mov.f32 	%f6, %f5;
	ld.param.f32 	%f7, [__cudaparmf1__Z18UnpremultiplyPixel8PixelRGB+12];
	mov.f32 	%f8, %f7;
	.loc	4	208	0
	cvt.ftz.sat.f32.f32 	%f9, %f8;
	mov.f32 	%f10, %f9;
	mov.f32 	%f11, 0fb70637bd;    	// -8e-006
	add.ftz.f32 	%f12, %f9, %f11;
	mov.f32 	%f13, 0f00000000;    	// 0
	setp.le.ftz.f32 	%p1, %f12, %f13;
	@%p1 bra 	$Lt_28_1282;
	.loc	4	213	0
	rcp.approx.ftz.f32 	%f14, %f9;
	mul.ftz.f32 	%f15, %f14, %f6;
	.loc	4	214	0
	mul.ftz.f32 	%f16, %f14, %f4;
	.loc	4	215	0
	mul.ftz.f32 	%f17, %f14, %f2;
	bra.uni 	$Lt_28_1026;
$Lt_28_1282:
	.loc	4	219	0
	mov.f32 	%f15, 0f00000000;    	// 0
	mov.f32 	%f16, 0f00000000;    	// 0
	mov.f32 	%f17, 0f00000000;    	// 0
	mov.f32 	%f10, 0f00000000;    	// 0
$Lt_28_1026:
	.loc	4	224	0
	mov.f32 	%f18, %f17;
	st.param.f32 	[__cudaretf__Z18UnpremultiplyPixel8PixelRGB+0], %f18;
	mov.f32 	%f19, %f16;
	st.param.f32 	[__cudaretf__Z18UnpremultiplyPixel8PixelRGB+4], %f19;
	mov.f32 	%f20, %f15;
	st.param.f32 	[__cudaretf__Z18UnpremultiplyPixel8PixelRGB+8], %f20;
	mov.f32 	%f21, %f10;
	st.param.f32 	[__cudaretf__Z18UnpremultiplyPixel8PixelRGB+12], %f21;
	ret;
$LDWend__Z18UnpremultiplyPixel8PixelRGB:
	} // _Z18UnpremultiplyPixel8PixelRGB

	.visible .func (.param .f32 __cudaretf__Z13ToLinearColorf) _Z13ToLinearColorf (.param .f32 __cudaparmf1__Z13ToLinearColorf)
	{
	.reg .f32 %f<15>;
	.reg .pred %p<3>;
	.loc	4	231	0
$LDWbegin__Z13ToLinearColorf:
	ld.param.f32 	%f1, [__cudaparmf1__Z13ToLinearColorf];
	mov.f32 	%f2, %f1;
	mov.f32 	%f3, 0f00000000;     	// 0
	setp.lt.ftz.f32 	%p1, %f2, %f3;
	@!%p1 bra 	$Lt_29_1026;
	.loc	4	234	0
	neg.ftz.f32 	%f4, %f2;
	lg2.approx.ftz.f32 	%f5, %f4;
	mov.f32 	%f6, 0f400ccccd;     	// 2.2
	mul.ftz.f32 	%f7, %f5, %f6;
	ex2.approx.ftz.f32 	%f8, %f7;
	neg.ftz.f32 	%f9, %f8;
	bra.uni 	$LBB4__Z13ToLinearColorf;
$Lt_29_1026:
	.loc	4	236	0
	lg2.approx.ftz.f32 	%f10, %f2;
	mov.f32 	%f11, 0f400ccccd;    	// 2.2
	mul.ftz.f32 	%f12, %f10, %f11;
	ex2.approx.ftz.f32 	%f9, %f12;
$LBB4__Z13ToLinearColorf:
	mov.f32 	%f13, %f9;
	st.param.f32 	[__cudaretf__Z13ToLinearColorf], %f13;
	ret;
$LDWend__Z13ToLinearColorf:
	} // _Z13ToLinearColorf

	.visible .func (.param .f32 __cudaretf__Z15FromLinearColorf) _Z15FromLinearColorf (.param .f32 __cudaparmf1__Z15FromLinearColorf)
	{
	.reg .f32 %f<15>;
	.reg .pred %p<3>;
	.loc	4	239	0
$LDWbegin__Z15FromLinearColorf:
	ld.param.f32 	%f1, [__cudaparmf1__Z15FromLinearColorf];
	mov.f32 	%f2, %f1;
	mov.f32 	%f3, 0f00000000;     	// 0
	setp.lt.ftz.f32 	%p1, %f2, %f3;
	@!%p1 bra 	$Lt_30_1026;
	.loc	4	242	0
	neg.ftz.f32 	%f4, %f2;
	lg2.approx.ftz.f32 	%f5, %f4;
	mov.f32 	%f6, 0f3ee8ba2e;     	// 0.454545
	mul.ftz.f32 	%f7, %f5, %f6;
	ex2.approx.ftz.f32 	%f8, %f7;
	neg.ftz.f32 	%f9, %f8;
	bra.uni 	$LBB4__Z15FromLinearColorf;
$Lt_30_1026:
	.loc	4	244	0
	lg2.approx.ftz.f32 	%f10, %f2;
	mov.f32 	%f11, 0f3ee8ba2e;    	// 0.454545
	mul.ftz.f32 	%f12, %f10, %f11;
	ex2.approx.ftz.f32 	%f9, %f12;
$LBB4__Z15FromLinearColorf:
	mov.f32 	%f13, %f9;
	st.param.f32 	[__cudaretf__Z15FromLinearColorf], %f13;
	ret;
$LDWend__Z15FromLinearColorf:
	} // _Z15FromLinearColorf

	.visible .func (.param .align 16 .b8 __cudaretf__Z25PremultiplyLinearizePixel8PixelRGB[16]) _Z25PremultiplyLinearizePixel8PixelRGB (.param .align 16 .b8 __cudaparmf1__Z25PremultiplyLinearizePixel8PixelRGB[16])
	{
	.reg .f32 %f<47>;
	.reg .pred %p<5>;
	.loc	4	252	0
$LDWbegin__Z25PremultiplyLinearizePixel8PixelRGB:
	ld.param.f32 	%f1, [__cudaparmf1__Z25PremultiplyLinearizePixel8PixelRGB+0];
	mov.f32 	%f2, %f1;
	ld.param.f32 	%f3, [__cudaparmf1__Z25PremultiplyLinearizePixel8PixelRGB+4];
	mov.f32 	%f4, %f3;
	ld.param.f32 	%f5, [__cudaparmf1__Z25PremultiplyLinearizePixel8PixelRGB+8];
	mov.f32 	%f6, %f5;
	ld.param.f32 	%f7, [__cudaparmf1__Z25PremultiplyLinearizePixel8PixelRGB+12];
	mov.f32 	%f8, %f7;
	.loc	4	254	0
	cvt.ftz.sat.f32.f32 	%f9, %f8;
	.loc	4	255	0
	mov.f32 	%f10, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p1, %f2, %f10;
	@!%p1 bra 	$Lt_31_4098;
	.loc	4	234	0
	neg.ftz.f32 	%f11, %f2;
	lg2.approx.ftz.f32 	%f12, %f11;
	mov.f32 	%f13, 0f400ccccd;    	// 2.2
	mul.ftz.f32 	%f14, %f12, %f13;
	ex2.approx.ftz.f32 	%f15, %f14;
	neg.ftz.f32 	%f16, %f15;
	bra.uni 	$LDWendi___log2f_209_5;
$Lt_31_4098:
	.loc	4	236	0
	lg2.approx.ftz.f32 	%f17, %f2;
	mov.f32 	%f18, 0f400ccccd;    	// 2.2
	mul.ftz.f32 	%f19, %f17, %f18;
	ex2.approx.ftz.f32 	%f16, %f19;
$LDWendi___log2f_209_5:
	.loc	4	256	0
	mov.f32 	%f20, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p2, %f4, %f20;
	@!%p2 bra 	$Lt_31_4610;
	.loc	4	234	0
	neg.ftz.f32 	%f21, %f4;
	lg2.approx.ftz.f32 	%f22, %f21;
	mov.f32 	%f23, 0f400ccccd;    	// 2.2
	mul.ftz.f32 	%f24, %f22, %f23;
	ex2.approx.ftz.f32 	%f25, %f24;
	neg.ftz.f32 	%f26, %f25;
	bra.uni 	$LDWendi___log2f_209_3;
$Lt_31_4610:
	.loc	4	236	0
	lg2.approx.ftz.f32 	%f27, %f4;
	mov.f32 	%f28, 0f400ccccd;    	// 2.2
	mul.ftz.f32 	%f29, %f27, %f28;
	ex2.approx.ftz.f32 	%f26, %f29;
$LDWendi___log2f_209_3:
	.loc	4	257	0
	mov.f32 	%f30, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p3, %f6, %f30;
	@!%p3 bra 	$Lt_31_5122;
	.loc	4	234	0
	neg.ftz.f32 	%f31, %f6;
	lg2.approx.ftz.f32 	%f32, %f31;
	mov.f32 	%f33, 0f400ccccd;    	// 2.2
	mul.ftz.f32 	%f34, %f32, %f33;
	ex2.approx.ftz.f32 	%f35, %f34;
	neg.ftz.f32 	%f36, %f35;
	bra.uni 	$LDWendi___log2f_209_1;
$Lt_31_5122:
	.loc	4	236	0
	lg2.approx.ftz.f32 	%f37, %f6;
	mov.f32 	%f38, 0f400ccccd;    	// 2.2
	mul.ftz.f32 	%f39, %f37, %f38;
	ex2.approx.ftz.f32 	%f36, %f39;
$LDWendi___log2f_209_1:
	.loc	4	259	0
	mul.ftz.f32 	%f40, %f36, %f9;
	mul.ftz.f32 	%f41, %f26, %f9;
	mul.ftz.f32 	%f42, %f16, %f9;
	st.param.f32 	[__cudaretf__Z25PremultiplyLinearizePixel8PixelRGB+0], %f42;
	mov.f32 	%f43, %f41;
	st.param.f32 	[__cudaretf__Z25PremultiplyLinearizePixel8PixelRGB+4], %f43;
	mov.f32 	%f44, %f40;
	st.param.f32 	[__cudaretf__Z25PremultiplyLinearizePixel8PixelRGB+8], %f44;
	mov.f32 	%f45, %f9;
	st.param.f32 	[__cudaretf__Z25PremultiplyLinearizePixel8PixelRGB+12], %f45;
	ret;
$LDWend__Z25PremultiplyLinearizePixel8PixelRGB:
	} // _Z25PremultiplyLinearizePixel8PixelRGB

	.visible .func (.param .align 16 .b8 __cudaretf__Z29UnpremultiplyUnlinearizePixel8PixelRGB[16]) _Z29UnpremultiplyUnlinearizePixel8PixelRGB (.param .align 16 .b8 __cudaparmf1__Z29UnpremultiplyUnlinearizePixel8PixelRGB[16])
	{
	.reg .f32 %f<53>;
	.reg .pred %p<6>;
	.loc	4	263	0
$LDWbegin__Z29UnpremultiplyUnlinearizePixel8PixelRGB:
	ld.param.f32 	%f1, [__cudaparmf1__Z29UnpremultiplyUnlinearizePixel8PixelRGB+0];
	mov.f32 	%f2, %f1;
	ld.param.f32 	%f3, [__cudaparmf1__Z29UnpremultiplyUnlinearizePixel8PixelRGB+4];
	mov.f32 	%f4, %f3;
	ld.param.f32 	%f5, [__cudaparmf1__Z29UnpremultiplyUnlinearizePixel8PixelRGB+8];
	mov.f32 	%f6, %f5;
	ld.param.f32 	%f7, [__cudaparmf1__Z29UnpremultiplyUnlinearizePixel8PixelRGB+12];
	mov.f32 	%f8, %f7;
	.loc	4	208	0
	cvt.ftz.sat.f32.f32 	%f9, %f8;
	mov.f32 	%f10, %f9;
	mov.f32 	%f11, 0fb70637bd;    	// -8e-006
	add.ftz.f32 	%f12, %f9, %f11;
	mov.f32 	%f13, 0f00000000;    	// 0
	setp.le.ftz.f32 	%p1, %f12, %f13;
	@%p1 bra 	$Lt_32_5122;
	.loc	4	213	0
	rcp.approx.ftz.f32 	%f14, %f9;
	mul.ftz.f32 	%f15, %f14, %f6;
	.loc	4	214	0
	mul.ftz.f32 	%f16, %f14, %f4;
	.loc	4	215	0
	mul.ftz.f32 	%f17, %f14, %f2;
	bra.uni 	$Lt_32_4866;
$Lt_32_5122:
	.loc	4	219	0
	mov.f32 	%f15, 0f00000000;    	// 0
	mov.f32 	%f16, 0f00000000;    	// 0
	mov.f32 	%f17, 0f00000000;    	// 0
	mov.f32 	%f10, 0f00000000;    	// 0
$Lt_32_4866:
	.loc	4	266	0
	mov.f32 	%f18, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p2, %f17, %f18;
	@!%p2 bra 	$Lt_32_5378;
	.loc	4	242	0
	neg.ftz.f32 	%f19, %f17;
	lg2.approx.ftz.f32 	%f20, %f19;
	mov.f32 	%f21, 0f3ee8ba2e;    	// 0.454545
	mul.ftz.f32 	%f22, %f20, %f21;
	ex2.approx.ftz.f32 	%f23, %f22;
	neg.ftz.f32 	%f24, %f23;
	bra.uni 	$LDWendi___log2f_210_5;
$Lt_32_5378:
	.loc	4	244	0
	lg2.approx.ftz.f32 	%f25, %f17;
	mov.f32 	%f26, 0f3ee8ba2e;    	// 0.454545
	mul.ftz.f32 	%f27, %f25, %f26;
	ex2.approx.ftz.f32 	%f24, %f27;
$LDWendi___log2f_210_5:
	.loc	4	267	0
	mov.f32 	%f28, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p3, %f16, %f28;
	@!%p3 bra 	$Lt_32_5890;
	.loc	4	242	0
	neg.ftz.f32 	%f29, %f16;
	lg2.approx.ftz.f32 	%f30, %f29;
	mov.f32 	%f31, 0f3ee8ba2e;    	// 0.454545
	mul.ftz.f32 	%f32, %f30, %f31;
	ex2.approx.ftz.f32 	%f33, %f32;
	neg.ftz.f32 	%f34, %f33;
	bra.uni 	$LDWendi___log2f_210_3;
$Lt_32_5890:
	.loc	4	244	0
	lg2.approx.ftz.f32 	%f35, %f16;
	mov.f32 	%f36, 0f3ee8ba2e;    	// 0.454545
	mul.ftz.f32 	%f37, %f35, %f36;
	ex2.approx.ftz.f32 	%f34, %f37;
$LDWendi___log2f_210_3:
	.loc	4	268	0
	mov.f32 	%f38, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p4, %f15, %f38;
	@!%p4 bra 	$Lt_32_6402;
	.loc	4	242	0
	neg.ftz.f32 	%f39, %f15;
	lg2.approx.ftz.f32 	%f40, %f39;
	mov.f32 	%f41, 0f3ee8ba2e;    	// 0.454545
	mul.ftz.f32 	%f42, %f40, %f41;
	ex2.approx.ftz.f32 	%f43, %f42;
	neg.ftz.f32 	%f44, %f43;
	bra.uni 	$LDWendi___log2f_210_1;
$Lt_32_6402:
	.loc	4	244	0
	lg2.approx.ftz.f32 	%f45, %f15;
	mov.f32 	%f46, 0f3ee8ba2e;    	// 0.454545
	mul.ftz.f32 	%f47, %f45, %f46;
	ex2.approx.ftz.f32 	%f44, %f47;
$LDWendi___log2f_210_1:
	.loc	4	269	0
	mov.f32 	%f48, %f24;
	st.param.f32 	[__cudaretf__Z29UnpremultiplyUnlinearizePixel8PixelRGB+0], %f48;
	mov.f32 	%f49, %f34;
	st.param.f32 	[__cudaretf__Z29UnpremultiplyUnlinearizePixel8PixelRGB+4], %f49;
	mov.f32 	%f50, %f44;
	st.param.f32 	[__cudaretf__Z29UnpremultiplyUnlinearizePixel8PixelRGB+8], %f50;
	mov.f32 	%f51, %f10;
	st.param.f32 	[__cudaretf__Z29UnpremultiplyUnlinearizePixel8PixelRGB+12], %f51;
	ret;
$LDWend__Z29UnpremultiplyUnlinearizePixel8PixelRGB:
	} // _Z29UnpremultiplyUnlinearizePixel8PixelRGB

	.visible .func (.param .align 16 .b8 __cudaretf__Z20PremultiplyLinearize6float4[16]) _Z20PremultiplyLinearize6float4 (.param .align 16 .b8 __cudaparmf1__Z20PremultiplyLinearize6float4[16])
	{
	.reg .f32 %f<47>;
	.reg .pred %p<5>;
	.loc	4	277	0
$LDWbegin__Z20PremultiplyLinearize6float4:
	ld.param.f32 	%f1, [__cudaparmf1__Z20PremultiplyLinearize6float4+0];
	mov.f32 	%f2, %f1;
	ld.param.f32 	%f3, [__cudaparmf1__Z20PremultiplyLinearize6float4+4];
	mov.f32 	%f4, %f3;
	ld.param.f32 	%f5, [__cudaparmf1__Z20PremultiplyLinearize6float4+8];
	mov.f32 	%f6, %f5;
	ld.param.f32 	%f7, [__cudaparmf1__Z20PremultiplyLinearize6float4+12];
	mov.f32 	%f8, %f7;
	.loc	4	254	0
	cvt.ftz.sat.f32.f32 	%f9, %f8;
	.loc	4	255	0
	mov.f32 	%f10, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p1, %f2, %f10;
	@!%p1 bra 	$Lt_33_4098;
	.loc	4	234	0
	neg.ftz.f32 	%f11, %f2;
	lg2.approx.ftz.f32 	%f12, %f11;
	mov.f32 	%f13, 0f400ccccd;    	// 2.2
	mul.ftz.f32 	%f14, %f12, %f13;
	ex2.approx.ftz.f32 	%f15, %f14;
	neg.ftz.f32 	%f16, %f15;
	bra.uni 	$LDWendi___log2f_211_5;
$Lt_33_4098:
	.loc	4	236	0
	lg2.approx.ftz.f32 	%f17, %f2;
	mov.f32 	%f18, 0f400ccccd;    	// 2.2
	mul.ftz.f32 	%f19, %f17, %f18;
	ex2.approx.ftz.f32 	%f16, %f19;
$LDWendi___log2f_211_5:
	.loc	4	256	0
	mov.f32 	%f20, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p2, %f4, %f20;
	@!%p2 bra 	$Lt_33_4610;
	.loc	4	234	0
	neg.ftz.f32 	%f21, %f4;
	lg2.approx.ftz.f32 	%f22, %f21;
	mov.f32 	%f23, 0f400ccccd;    	// 2.2
	mul.ftz.f32 	%f24, %f22, %f23;
	ex2.approx.ftz.f32 	%f25, %f24;
	neg.ftz.f32 	%f26, %f25;
	bra.uni 	$LDWendi___log2f_211_3;
$Lt_33_4610:
	.loc	4	236	0
	lg2.approx.ftz.f32 	%f27, %f4;
	mov.f32 	%f28, 0f400ccccd;    	// 2.2
	mul.ftz.f32 	%f29, %f27, %f28;
	ex2.approx.ftz.f32 	%f26, %f29;
$LDWendi___log2f_211_3:
	.loc	4	257	0
	mov.f32 	%f30, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p3, %f6, %f30;
	@!%p3 bra 	$Lt_33_5122;
	.loc	4	234	0
	neg.ftz.f32 	%f31, %f6;
	lg2.approx.ftz.f32 	%f32, %f31;
	mov.f32 	%f33, 0f400ccccd;    	// 2.2
	mul.ftz.f32 	%f34, %f32, %f33;
	ex2.approx.ftz.f32 	%f35, %f34;
	neg.ftz.f32 	%f36, %f35;
	bra.uni 	$LDWendi___log2f_211_1;
$Lt_33_5122:
	.loc	4	236	0
	lg2.approx.ftz.f32 	%f37, %f6;
	mov.f32 	%f38, 0f400ccccd;    	// 2.2
	mul.ftz.f32 	%f39, %f37, %f38;
	ex2.approx.ftz.f32 	%f36, %f39;
$LDWendi___log2f_211_1:
	.loc	4	259	0
	mul.ftz.f32 	%f40, %f36, %f9;
	mul.ftz.f32 	%f41, %f26, %f9;
	.loc	4	278	0
	mul.ftz.f32 	%f42, %f16, %f9;
	st.param.f32 	[__cudaretf__Z20PremultiplyLinearize6float4+0], %f42;
	mov.f32 	%f43, %f41;
	st.param.f32 	[__cudaretf__Z20PremultiplyLinearize6float4+4], %f43;
	mov.f32 	%f44, %f40;
	st.param.f32 	[__cudaretf__Z20PremultiplyLinearize6float4+8], %f44;
	mov.f32 	%f45, %f9;
	st.param.f32 	[__cudaretf__Z20PremultiplyLinearize6float4+12], %f45;
	ret;
$LDWend__Z20PremultiplyLinearize6float4:
	} // _Z20PremultiplyLinearize6float4

	.visible .func (.param .align 16 .b8 __cudaretf__Z24UnpremultiplyUnlinearize6float4[16]) _Z24UnpremultiplyUnlinearize6float4 (.param .align 16 .b8 __cudaparmf1__Z24UnpremultiplyUnlinearize6float4[16])
	{
	.reg .f32 %f<53>;
	.reg .pred %p<6>;
	.loc	4	284	0
$LDWbegin__Z24UnpremultiplyUnlinearize6float4:
	ld.param.f32 	%f1, [__cudaparmf1__Z24UnpremultiplyUnlinearize6float4+0];
	mov.f32 	%f2, %f1;
	ld.param.f32 	%f3, [__cudaparmf1__Z24UnpremultiplyUnlinearize6float4+4];
	mov.f32 	%f4, %f3;
	ld.param.f32 	%f5, [__cudaparmf1__Z24UnpremultiplyUnlinearize6float4+8];
	mov.f32 	%f6, %f5;
	ld.param.f32 	%f7, [__cudaparmf1__Z24UnpremultiplyUnlinearize6float4+12];
	mov.f32 	%f8, %f7;
	.loc	4	208	0
	cvt.ftz.sat.f32.f32 	%f9, %f8;
	mov.f32 	%f10, %f9;
	mov.f32 	%f11, 0fb70637bd;    	// -8e-006
	add.ftz.f32 	%f12, %f9, %f11;
	mov.f32 	%f13, 0f00000000;    	// 0
	setp.le.ftz.f32 	%p1, %f12, %f13;
	@%p1 bra 	$Lt_34_5122;
	.loc	4	213	0
	rcp.approx.ftz.f32 	%f14, %f9;
	mul.ftz.f32 	%f15, %f14, %f6;
	.loc	4	214	0
	mul.ftz.f32 	%f16, %f14, %f4;
	.loc	4	215	0
	mul.ftz.f32 	%f17, %f14, %f2;
	bra.uni 	$Lt_34_4866;
$Lt_34_5122:
	.loc	4	219	0
	mov.f32 	%f15, 0f00000000;    	// 0
	mov.f32 	%f16, 0f00000000;    	// 0
	mov.f32 	%f17, 0f00000000;    	// 0
	mov.f32 	%f10, 0f00000000;    	// 0
$Lt_34_4866:
	.loc	4	266	0
	mov.f32 	%f18, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p2, %f17, %f18;
	@!%p2 bra 	$Lt_34_5378;
	.loc	4	242	0
	neg.ftz.f32 	%f19, %f17;
	lg2.approx.ftz.f32 	%f20, %f19;
	mov.f32 	%f21, 0f3ee8ba2e;    	// 0.454545
	mul.ftz.f32 	%f22, %f20, %f21;
	ex2.approx.ftz.f32 	%f23, %f22;
	neg.ftz.f32 	%f24, %f23;
	bra.uni 	$LDWendi___log2f_212_5;
$Lt_34_5378:
	.loc	4	244	0
	lg2.approx.ftz.f32 	%f25, %f17;
	mov.f32 	%f26, 0f3ee8ba2e;    	// 0.454545
	mul.ftz.f32 	%f27, %f25, %f26;
	ex2.approx.ftz.f32 	%f24, %f27;
$LDWendi___log2f_212_5:
	.loc	4	267	0
	mov.f32 	%f28, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p3, %f16, %f28;
	@!%p3 bra 	$Lt_34_5890;
	.loc	4	242	0
	neg.ftz.f32 	%f29, %f16;
	lg2.approx.ftz.f32 	%f30, %f29;
	mov.f32 	%f31, 0f3ee8ba2e;    	// 0.454545
	mul.ftz.f32 	%f32, %f30, %f31;
	ex2.approx.ftz.f32 	%f33, %f32;
	neg.ftz.f32 	%f34, %f33;
	bra.uni 	$LDWendi___log2f_212_3;
$Lt_34_5890:
	.loc	4	244	0
	lg2.approx.ftz.f32 	%f35, %f16;
	mov.f32 	%f36, 0f3ee8ba2e;    	// 0.454545
	mul.ftz.f32 	%f37, %f35, %f36;
	ex2.approx.ftz.f32 	%f34, %f37;
$LDWendi___log2f_212_3:
	.loc	4	268	0
	mov.f32 	%f38, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p4, %f15, %f38;
	@!%p4 bra 	$Lt_34_6402;
	.loc	4	242	0
	neg.ftz.f32 	%f39, %f15;
	lg2.approx.ftz.f32 	%f40, %f39;
	mov.f32 	%f41, 0f3ee8ba2e;    	// 0.454545
	mul.ftz.f32 	%f42, %f40, %f41;
	ex2.approx.ftz.f32 	%f43, %f42;
	neg.ftz.f32 	%f44, %f43;
	bra.uni 	$LDWendi___log2f_212_1;
$Lt_34_6402:
	.loc	4	244	0
	lg2.approx.ftz.f32 	%f45, %f15;
	mov.f32 	%f46, 0f3ee8ba2e;    	// 0.454545
	mul.ftz.f32 	%f47, %f45, %f46;
	ex2.approx.ftz.f32 	%f44, %f47;
$LDWendi___log2f_212_1:
	.loc	4	285	0
	mov.f32 	%f48, %f24;
	st.param.f32 	[__cudaretf__Z24UnpremultiplyUnlinearize6float4+0], %f48;
	mov.f32 	%f49, %f34;
	st.param.f32 	[__cudaretf__Z24UnpremultiplyUnlinearize6float4+4], %f49;
	mov.f32 	%f50, %f44;
	st.param.f32 	[__cudaretf__Z24UnpremultiplyUnlinearize6float4+8], %f50;
	mov.f32 	%f51, %f10;
	st.param.f32 	[__cudaretf__Z24UnpremultiplyUnlinearize6float4+12], %f51;
	ret;
$LDWend__Z24UnpremultiplyUnlinearize6float4:
	} // _Z24UnpremultiplyUnlinearize6float4

	.visible .func (.param .align 16 .b8 __cudaretf__Z23ConvertPixel_422_To_4446float414IR_PixelFormatS0_[16]) _Z23ConvertPixel_422_To_4446float414IR_PixelFormatS0_ (.param .align 16 .b8 __cudaparmf1__Z23ConvertPixel_422_To_4446float414IR_PixelFormatS0_[16], .param .s32 __cudaparmf2__Z23ConvertPixel_422_To_4446float414IR_PixelFormatS0_, .param .s32 __cudaparmf3__Z23ConvertPixel_422_To_4446float414IR_PixelFormatS0_)
	{
	.reg .u32 %r<194>;
	.reg .u64 %rd<3>;
	.reg .f32 %f<202>;
	.reg .pred %p<138>;
	.loc	22	34	0
$LDWbegin__Z23ConvertPixel_422_To_4446float414IR_PixelFormatS0_:
	ld.param.f32 	%f1, [__cudaparmf1__Z23ConvertPixel_422_To_4446float414IR_PixelFormatS0_+0];
	mov.f32 	%f2, %f1;
	mov.f32 	%f3, %f2;
	ld.param.f32 	%f4, [__cudaparmf1__Z23ConvertPixel_422_To_4446float414IR_PixelFormatS0_+4];
	mov.f32 	%f5, %f4;
	ld.param.f32 	%f6, [__cudaparmf1__Z23ConvertPixel_422_To_4446float414IR_PixelFormatS0_+8];
	mov.f32 	%f7, %f6;
	ld.param.f32 	%f8, [__cudaparmf1__Z23ConvertPixel_422_To_4446float414IR_PixelFormatS0_+12];
	mov.f32 	%f9, %f8;
	ld.param.u32 	%r1, [__cudaparmf2__Z23ConvertPixel_422_To_4446float414IR_PixelFormatS0_];
	mov.s32 	%r2, %r1;
	ld.param.u32 	%r3, [__cudaparmf3__Z23ConvertPixel_422_To_4446float414IR_PixelFormatS0_];
	mov.s32 	%r4, %r3;
	and.b32 	%r5, %r2, 4096;
	mov.u32 	%r6, 0;
	setp.eq.s32 	%p1, %r5, %r6;
	@%p1 bra 	$Lt_35_228866;
	.loc	23	86	0
	mov.f32 	%f3, %f5;
	.loc	23	87	0
	mov.f32 	%f5, %f2;
	.loc	22	38	0
	mov.f32 	%f7, %f9;
$Lt_35_228866:
	.loc	20	57	0
	and.b32 	%r7, %r2, 1536;
	and.b32 	%r8, %r2, 448;
	mov.s32 	%r9, 512;
	setp.eq.s32 	%p2, %r7, %r9;
	mov.s32 	%r10, 2565;
	mov.s32 	%r11, 2053;
	selp.s32 	%r12, %r10, %r11, %p2;
	or.b32 	%r13, %r12, %r8;
	and.b32 	%r14, %r13, 448;
	mov.s32 	%r15, 0;
	setp.eq.s32 	%p3, %r14, %r15;
	@!%p3 bra 	$Lt_35_144642;
	.loc	20	59	0
	mov.f32 	%f10, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_77;
$Lt_35_144642:
	.loc	20	61	0
	mov.u32 	%r16, 64;
	setp.ne.s32 	%p4, %r14, %r16;
	@%p4 bra 	$Lt_35_144898;
	.loc	20	63	0
	mov.f32 	%f10, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_77;
$Lt_35_144898:
	.loc	20	65	0
	mov.u32 	%r17, 128;
	setp.ne.s32 	%p5, %r14, %r17;
	@%p5 bra 	$Lt_35_145154;
	.loc	20	68	0
	mov.f32 	%f10, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_77;
$Lt_35_145154:
	.loc	20	70	0
	mov.u32 	%r18, 192;
	setp.ne.s32 	%p6, %r14, %r18;
	@%p6 bra 	$Lt_35_145410;
	.loc	20	72	0
	mov.f32 	%f10, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_77;
$Lt_35_145410:
	.loc	20	76	0
	mov.f32 	%f10, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_213_77:
	.loc	20	469	0
	mov.f32 	%f11, %f7;
	mov.f32 	%f12, %f11;
	mov.f32 	%f13, %f3;
	mov.f32 	%f14, %f13;
	mov.f32 	%f15, %f5;
	mov.f32 	%f16, %f10;
	and.b32 	%r19, %r13, 4096;
	mov.u32 	%r20, 0;
	setp.ne.s32 	%p7, %r19, %r20;
	@%p7 bra 	$Lt_35_229378;
	.loc	20	473	0
	mov.f32 	%f12, %f16;
	mov.f32 	%f14, %f15;
	mov.f32 	%f15, %f13;
	mov.f32 	%f16, %f11;
$Lt_35_229378:
	.loc	20	476	0
	and.b32 	%r21, %r4, 448;
	mov.s32 	%r22, %r4;
	mov.s32 	%r23, 256;
	setp.ne.s32 	%p8, %r21, %r23;
	and.b32 	%r24, %r22, 1;
	selp.s32 	%r25, 1, 0, %p8;
	and.b32 	%r26, %r24, %r25;
	mov.s32 	%r27, %r13;
	mov.s32 	%r28, 256;
	setp.ne.s32 	%p9, %r14, %r28;
	and.b32 	%r29, %r27, 1;
	selp.s32 	%r30, 1, 0, %p9;
	and.b32 	%r31, %r29, %r30;
	mov.u32 	%r32, 0;
	setp.eq.s32 	%p10, %r31, %r32;
	@%p10 bra 	$Lt_35_256514;
	mov.u32 	%r33, 0;
	setp.ne.s32 	%p11, %r26, %r33;
	@%p11 bra 	$Lt_35_256514;
	.loc	20	57	0
	@!%p3 bra 	$Lt_35_145922;
	.loc	20	59	0
	mov.f32 	%f17, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_75;
$Lt_35_145922:
	.loc	20	61	0
	mov.u32 	%r34, 64;
	setp.ne.s32 	%p12, %r14, %r34;
	@%p12 bra 	$Lt_35_146178;
	.loc	20	63	0
	mov.f32 	%f17, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_75;
$Lt_35_146178:
	.loc	20	65	0
	mov.u32 	%r35, 128;
	setp.ne.s32 	%p13, %r14, %r35;
	@%p13 bra 	$Lt_35_146434;
	.loc	20	68	0
	mov.f32 	%f17, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_75;
$Lt_35_146434:
	.loc	20	70	0
	mov.u32 	%r36, 192;
	setp.ne.s32 	%p14, %r14, %r36;
	@%p14 bra 	$Lt_35_146690;
	.loc	20	72	0
	mov.f32 	%f17, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_75;
$Lt_35_146690:
	.loc	20	76	0
	mov.f32 	%f17, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_213_75:
	.loc	20	118	0
	and.b32 	%r37, %r13, 2048;
	mov.s32 	%r38, 0;
	setp.ne.s32 	%p15, %r37, %r38;
	@!%p15 bra 	$Lt_35_230146;
	.loc	20	100	0
	ld.const.f32 	%f18, [kYCbCrOffset+0];
	bra.uni 	$Lt_35_229890;
$Lt_35_230146:
	ld.const.f32 	%f18, [kYCbCrFullRangeOffset+0];
$Lt_35_229890:
	.loc	20	118	0
	@!%p15 bra 	$Lt_35_230658;
	.loc	20	100	0
	ld.const.f32 	%f19, [kYCbCrOffset+4];
	bra.uni 	$Lt_35_230402;
$Lt_35_230658:
	ld.const.f32 	%f19, [kYCbCrFullRangeOffset+4];
$Lt_35_230402:
	.loc	20	118	0
	@!%p15 bra 	$Lt_35_231170;
	.loc	20	100	0
	ld.const.f32 	%f20, [kYCbCrOffset+8];
	bra.uni 	$Lt_35_230914;
$Lt_35_231170:
	ld.const.f32 	%f20, [kYCbCrFullRangeOffset+8];
$Lt_35_230914:
	.loc	20	478	0
	mov.f32 	%f21, 0f437f0000;    	// 255
	div.approx.ftz.f32 	%f22, %f17, %f21;
	mul.ftz.f32 	%f23, %f22, %f18;
	sub.ftz.f32 	%f14, %f14, %f23;
	mul.ftz.f32 	%f24, %f22, %f19;
	sub.ftz.f32 	%f15, %f15, %f24;
	mul.ftz.f32 	%f25, %f22, %f20;
	sub.ftz.f32 	%f16, %f16, %f25;
$Lt_35_256514:
$Lt_35_29698:
	.loc	20	481	0
	and.b32 	%r39, %r4, 2;
	and.b32 	%r40, %r13, 2;
	mov.u32 	%r41, 0;
	setp.eq.s32 	%p16, %r40, %r41;
	@%p16 bra 	$Lt_35_257026;
	mov.u32 	%r42, 0;
	setp.ne.s32 	%p17, %r39, %r42;
	@%p17 bra 	$Lt_35_257026;
	.loc	20	483	0
	mov.f32 	%f26, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p18, %f14, %f26;
	@!%p18 bra 	$Lt_35_231426;
	.loc	20	372	0
	neg.ftz.f32 	%f27, %f14;
	lg2.approx.ftz.f32 	%f28, %f27;
	mov.f32 	%f29, 0f3ee66666;    	// 0.45
	mul.ftz.f32 	%f30, %f28, %f29;
	ex2.approx.ftz.f32 	%f31, %f30;
	neg.ftz.f32 	%f32, %f31;
	bra.uni 	$LDWendi___log2f_213_71;
$Lt_35_231426:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f33, %f14;
	mov.f32 	%f34, 0f3ee66666;    	// 0.45
	mul.ftz.f32 	%f35, %f33, %f34;
	ex2.approx.ftz.f32 	%f32, %f35;
$LDWendi___log2f_213_71:
	.loc	20	483	0
	mov.f32 	%f36, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p19, %f15, %f36;
	@!%p19 bra 	$Lt_35_231938;
	.loc	20	372	0
	neg.ftz.f32 	%f37, %f15;
	lg2.approx.ftz.f32 	%f38, %f37;
	mov.f32 	%f39, 0f3ee66666;    	// 0.45
	mul.ftz.f32 	%f40, %f38, %f39;
	ex2.approx.ftz.f32 	%f41, %f40;
	neg.ftz.f32 	%f42, %f41;
	bra.uni 	$LDWendi___log2f_213_69;
$Lt_35_231938:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f43, %f15;
	mov.f32 	%f44, 0f3ee66666;    	// 0.45
	mul.ftz.f32 	%f45, %f43, %f44;
	ex2.approx.ftz.f32 	%f42, %f45;
$LDWendi___log2f_213_69:
	.loc	20	483	0
	mov.f32 	%f46, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p20, %f16, %f46;
	@!%p20 bra 	$Lt_35_232450;
	.loc	20	372	0
	neg.ftz.f32 	%f47, %f16;
	lg2.approx.ftz.f32 	%f48, %f47;
	mov.f32 	%f49, 0f3ee66666;    	// 0.45
	mul.ftz.f32 	%f50, %f48, %f49;
	ex2.approx.ftz.f32 	%f51, %f50;
	neg.ftz.f32 	%f52, %f51;
	bra.uni 	$LDWendi___log2f_213_67;
$Lt_35_232450:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f53, %f16;
	mov.f32 	%f54, 0f3ee66666;    	// 0.45
	mul.ftz.f32 	%f55, %f53, %f54;
	ex2.approx.ftz.f32 	%f52, %f55;
$LDWendi___log2f_213_67:
	.loc	20	483	0
	mov.f32 	%f14, %f32;
	mov.f32 	%f15, %f42;
	mov.f32 	%f16, %f52;
$Lt_35_257026:
$Lt_35_33538:
	.loc	20	486	0
	and.b32 	%r43, %r4, 1;
	and.b32 	%r44, %r4, 1536;
	and.b32 	%r45, %r13, 1;
	and.b32 	%r46, %r13, 1536;
	set.ne.u32.s32 	%r47, %r43, %r45;
	neg.s32 	%r48, %r47;
	set.ne.u32.s32 	%r49, %r44, %r46;
	neg.s32 	%r50, %r49;
	or.b32 	%r51, %r48, %r50;
	mov.u32 	%r52, 0;
	setp.ne.s32 	%p21, %r51, %r52;
	@%p21 bra 	$Lt_35_84226;
	setp.eq.s32 	%p22, %r26, %r31;
	@%p22 bra 	$Lt_35_84482;
$Lt_35_84226:
	.loc	20	490	0
	mov.u32 	%r53, 0;
	setp.ne.s32 	%p23, %r45, %r53;
	@%p23 bra 	$Lt_35_233218;
	mov.s32 	%r54, 256;
	setp.eq.s32 	%p24, %r21, %r54;
	mov.u32 	%r55, 256;
	setp.ne.s32 	%p25, %r14, %r55;
	@%p25 bra 	$Lt_35_233730;
	.loc	20	137	0
	mov.s32 	%r56, 512;
	setp.eq.s32 	%p26, %r44, %r56;
	@!%p24 bra 	$Lt_35_150018;
	.loc	20	139	0
	@!%p26 bra 	$Lt_35_150274;
	.loc	20	141	0
	mov.u64 	%rd1, kRGB32f_To_709YPbPr;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__213_65;
$Lt_35_150274:
	.loc	20	145	0
	mov.u64 	%rd1, kRGB32f_To_601YPbPr;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__213_65;
$Lt_35_150018:
	.loc	20	150	0
	@!%p26 bra 	$Lt_35_150530;
	.loc	20	152	0
	mov.u64 	%rd1, kRGB32f_To_709YCbCr;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__213_65;
$Lt_35_150530:
	.loc	20	154	0
	and.b32 	%r57, %r4, 2048;
	mov.u32 	%r58, 0;
	setp.ne.s32 	%p27, %r57, %r58;
	@%p27 bra 	$Lt_35_150786;
	.loc	20	156	0
	mov.u64 	%rd1, kRGB32f_To_601YCbCrFullRange;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__213_65;
$Lt_35_150786:
	.loc	20	160	0
	mov.u64 	%rd1, kRGB32f_To_601YCbCr;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__213_65;
$Lt_35_233730:
	@!%p24 bra 	$Lt_35_234242;
	bra.uni 	$Lt_35_232962;
$Lt_35_234242:
	.loc	20	179	0
	mov.u32 	%r59, 512;
	setp.ne.s32 	%p28, %r44, %r59;
	@%p28 bra 	$Lt_35_151554;
	.loc	20	181	0
	mov.u64 	%rd1, kRGB8u_To_709YCbCr;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__213_65;
$Lt_35_151554:
	.loc	20	183	0
	and.b32 	%r60, %r4, 2048;
	mov.u32 	%r61, 0;
	setp.ne.s32 	%p29, %r60, %r61;
	@%p29 bra 	$Lt_35_151810;
	.loc	20	185	0
	mov.u64 	%rd1, kRGB8u_To_601YCbCrFullRange;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__213_65;
$Lt_35_151810:
	.loc	20	189	0
	mov.u64 	%rd1, kRGB8u_To_601YCbCr;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__213_65;
$Lt_35_233218:
	mov.s32 	%r62, 0;
	setp.eq.s32 	%p30, %r43, %r62;
	mov.u32 	%r63, 512;
	setp.ne.s32 	%p31, %r46, %r63;
	@%p31 bra 	$Lt_35_234754;
	mov.s32 	%r64, 256;
	setp.eq.s32 	%p32, %r14, %r64;
	@!%p30 bra 	$Lt_35_235266;
	mov.s32 	%r65, 256;
	setp.eq.s32 	%p33, %r21, %r65;
	@!%p32 bra 	$Lt_35_235778;
	@!%p33 bra 	$Lt_35_232962;
	.loc	20	202	0
	mov.u64 	%rd1, k709YPbPr_To_RGB32f;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__213_65;
$Lt_35_235778:
	.loc	20	211	0
	@!%p33 bra 	$Lt_35_153090;
	.loc	20	213	0
	mov.u64 	%rd1, k709YCbCr_To_RGB32f;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__213_65;
$Lt_35_153090:
	.loc	20	217	0
	mov.u64 	%rd1, k709YCbCr_To_RGB8u;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__213_65;
$Lt_35_235266:
	@!%p32 bra 	$Lt_35_236802;
	bra.uni 	$Lt_35_232962;
$Lt_35_236802:
	mov.s32 	%r66, 256;
	set.eq.u32.s32 	%r67, %r21, %r66;
	neg.s32 	%r68, %r67;
	and.b32 	%r69, %r4, 2048;
	mov.s32 	%r70, 0;
	set.eq.u32.s32 	%r71, %r69, %r70;
	neg.s32 	%r72, %r71;
	or.b32 	%r73, %r68, %r72;
	mov.u32 	%r74, 0;
	setp.eq.s32 	%p34, %r73, %r74;
	@%p34 bra 	$Lt_35_237314;
	bra.uni 	$Lt_35_232962;
$Lt_35_237314:
	.loc	20	250	0
	mov.u64 	%rd1, k709YCbCr_To_601YCbCr;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__213_65;
$Lt_35_234754:
	and.b32 	%r75, %r13, 2048;
	mov.s32 	%r76, 0;
	setp.eq.s32 	%p35, %r75, %r76;
	@!%p35 bra 	$Lt_35_237826;
	@!%p30 bra 	$Lt_35_232962;
	.loc	20	259	0
	mov.u32 	%r77, 256;
	setp.ne.s32 	%p36, %r21, %r77;
	@%p36 bra 	$Lt_35_154626;
	.loc	20	261	0
	mov.u64 	%rd1, k601YCbCrFullRange_To_RGB32f;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__213_65;
$Lt_35_154626:
	.loc	20	265	0
	mov.u64 	%rd1, k601YCbCrFullRange_To_RGB8u;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__213_65;
$Lt_35_237826:
	mov.s32 	%r78, 256;
	setp.eq.s32 	%p32, %r14, %r78;
	@!%p30 bra 	$Lt_35_238850;
	mov.s32 	%r79, 256;
	setp.eq.s32 	%p37, %r21, %r79;
	@!%p32 bra 	$Lt_35_239362;
	@!%p37 bra 	$Lt_35_232962;
	.loc	20	302	0
	mov.u64 	%rd1, k601YPbPr_To_RGB32f;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__213_65;
$Lt_35_239362:
	.loc	20	311	0
	@!%p37 bra 	$Lt_35_156418;
	.loc	20	313	0
	mov.u64 	%rd1, k601YCbCr_To_RGB32f;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__213_65;
$Lt_35_156418:
	.loc	20	317	0
	mov.u64 	%rd1, k601YCbCr_To_RGB8u;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__213_65;
$Lt_35_238850:
	@!%p32 bra 	$Lt_35_240386;
	bra.uni 	$Lt_35_232962;
$Lt_35_240386:
	selp.s32 	%r80, 1, 0, %p35;
	mov.s32 	%r81, 256;
	set.eq.u32.s32 	%r82, %r21, %r81;
	neg.s32 	%r83, %r82;
	or.b32 	%r84, %r80, %r83;
	mov.u32 	%r85, 0;
	setp.eq.s32 	%p38, %r84, %r85;
	@%p38 bra 	$Lt_35_240898;
	bra.uni 	$Lt_35_232962;
$Lt_35_240898:
	.loc	20	350	0
	mov.u64 	%rd1, k601YCbCr_To_709YCbCr;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__213_65;
$Lt_35_232962:
	.loc	20	355	0
	mov.u64 	%rd1, 0;
$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__213_65:
	.loc	20	490	0
	ld.global.f32 	%f56, [%rd1+16];
	mul.ftz.f32 	%f57, %f56, %f15;
	ld.global.f32 	%f58, [%rd1+12];
	fma.rn.ftz.f32 	%f59, %f58, %f14, %f57;
	ld.global.f32 	%f60, [%rd1+20];
	fma.rn.ftz.f32 	%f61, %f60, %f16, %f59;
	ld.global.f32 	%f62, [%rd1+28];
	mul.ftz.f32 	%f63, %f62, %f15;
	ld.global.f32 	%f64, [%rd1+24];
	fma.rn.ftz.f32 	%f65, %f64, %f14, %f63;
	ld.global.f32 	%f66, [%rd1+32];
	fma.rn.ftz.f32 	%f67, %f66, %f16, %f65;
	ld.global.f32 	%f68, [%rd1+4];
	mul.ftz.f32 	%f69, %f68, %f15;
	ld.global.f32 	%f70, [%rd1+0];
	fma.rn.ftz.f32 	%f71, %f70, %f14, %f69;
	ld.global.f32 	%f72, [%rd1+8];
	fma.rn.ftz.f32 	%f14, %f72, %f16, %f71;
	mov.f32 	%f15, %f61;
	mov.f32 	%f16, %f67;
	setp.eq.s32 	%p39, %r21, %r14;
	@%p39 bra 	$Lt_35_241666;
	.loc	20	494	0
	mov.s32 	%r86, 256;
	setp.eq.s32 	%p32, %r14, %r86;
	@!%p32 bra 	$L_35_226306;
	mov.s32 	%r87, 0;
	setp.eq.s32 	%p40, %r21, %r87;
	@%p40 bra 	$L_35_226050;
$L_35_226306:
	@!%p3 bra 	$Lt_35_258306;
	mov.u32 	%r88, 256;
	setp.ne.s32 	%p41, %r21, %r88;
	@%p41 bra 	$Lt_35_258306;
	mov.s32 	%r89, 0;
	setp.eq.s32 	%p40, %r21, %r89;
$L_35_226050:
	.loc	20	57	0
	@!%p40 bra 	$Lt_35_157954;
	.loc	20	59	0
	mov.f32 	%f73, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_63;
$Lt_35_157954:
	.loc	20	61	0
	mov.u32 	%r90, 64;
	setp.ne.s32 	%p42, %r21, %r90;
	@%p42 bra 	$Lt_35_158210;
	.loc	20	63	0
	mov.f32 	%f73, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_63;
$Lt_35_158210:
	.loc	20	65	0
	mov.u32 	%r91, 128;
	setp.ne.s32 	%p43, %r21, %r91;
	@%p43 bra 	$Lt_35_158466;
	.loc	20	68	0
	mov.f32 	%f73, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_63;
$Lt_35_158466:
	.loc	20	70	0
	mov.u32 	%r92, 192;
	setp.ne.s32 	%p44, %r21, %r92;
	@%p44 bra 	$Lt_35_158722;
	.loc	20	72	0
	mov.f32 	%f73, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_63;
$Lt_35_158722:
	.loc	20	76	0
	mov.f32 	%f73, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_213_63:
	.loc	20	57	0
	@!%p3 bra 	$Lt_35_158978;
	.loc	20	59	0
	mov.f32 	%f74, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_61;
$Lt_35_158978:
	.loc	20	61	0
	mov.u32 	%r93, 64;
	setp.ne.s32 	%p45, %r14, %r93;
	@%p45 bra 	$Lt_35_159234;
	.loc	20	63	0
	mov.f32 	%f74, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_61;
$Lt_35_159234:
	.loc	20	65	0
	mov.u32 	%r94, 128;
	setp.ne.s32 	%p46, %r14, %r94;
	@%p46 bra 	$Lt_35_159490;
	.loc	20	68	0
	mov.f32 	%f74, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_61;
$Lt_35_159490:
	.loc	20	70	0
	mov.u32 	%r95, 192;
	setp.ne.s32 	%p47, %r14, %r95;
	@%p47 bra 	$Lt_35_159746;
	.loc	20	72	0
	mov.f32 	%f74, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_61;
$Lt_35_159746:
	.loc	20	76	0
	mov.f32 	%f74, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_213_61:
	.loc	20	498	0
	div.approx.ftz.f32 	%f75, %f73, %f74;
	mul.ftz.f32 	%f12, %f12, %f75;
	bra.uni 	$Lt_35_241666;
$Lt_35_258306:
$L_35_225794:
	.loc	20	500	0
	@!%p32 bra 	$L_35_227330;
	@%p8 bra 	$L_35_227074;
$L_35_227330:
	@!%p9 bra 	$Lt_35_259330;
	mov.u32 	%r96, 256;
	setp.ne.s32 	%p48, %r21, %r96;
	@%p48 bra 	$Lt_35_259330;
$L_35_227074:
	.loc	20	57	0
	mov.u32 	%r97, 0;
	setp.ne.s32 	%p49, %r21, %r97;
	@%p49 bra 	$Lt_35_160258;
	.loc	20	59	0
	mov.f32 	%f73, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_59;
$Lt_35_160258:
	.loc	20	61	0
	mov.u32 	%r98, 64;
	setp.ne.s32 	%p50, %r21, %r98;
	@%p50 bra 	$Lt_35_160514;
	.loc	20	63	0
	mov.f32 	%f73, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_59;
$Lt_35_160514:
	.loc	20	65	0
	mov.u32 	%r99, 128;
	setp.ne.s32 	%p51, %r21, %r99;
	@%p51 bra 	$Lt_35_160770;
	.loc	20	68	0
	mov.f32 	%f73, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_59;
$Lt_35_160770:
	.loc	20	70	0
	mov.u32 	%r100, 192;
	setp.ne.s32 	%p52, %r21, %r100;
	@%p52 bra 	$Lt_35_161026;
	.loc	20	72	0
	mov.f32 	%f73, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_59;
$Lt_35_161026:
	.loc	20	76	0
	mov.f32 	%f73, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_213_59:
	.loc	20	57	0
	@!%p3 bra 	$Lt_35_161282;
	.loc	20	59	0
	mov.f32 	%f74, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_57;
$Lt_35_161282:
	.loc	20	61	0
	mov.u32 	%r101, 64;
	setp.ne.s32 	%p53, %r14, %r101;
	@%p53 bra 	$Lt_35_161538;
	.loc	20	63	0
	mov.f32 	%f74, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_57;
$Lt_35_161538:
	.loc	20	65	0
	mov.u32 	%r102, 128;
	setp.ne.s32 	%p54, %r14, %r102;
	@%p54 bra 	$Lt_35_161794;
	.loc	20	68	0
	mov.f32 	%f74, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_57;
$Lt_35_161794:
	.loc	20	70	0
	mov.u32 	%r103, 192;
	setp.ne.s32 	%p55, %r14, %r103;
	@%p55 bra 	$Lt_35_162050;
	.loc	20	72	0
	mov.f32 	%f74, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_57;
$Lt_35_162050:
	.loc	20	76	0
	mov.f32 	%f74, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_213_57:
	.loc	20	504	0
	div.approx.ftz.f32 	%f76, %f73, %f74;
	mul.ftz.f32 	%f12, %f12, %f76;
	.loc	20	57	0
	@!%p3 bra 	$Lt_35_163330;
	.loc	20	59	0
	mov.f32 	%f74, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_53;
$Lt_35_163330:
	.loc	20	61	0
	mov.u32 	%r104, 64;
	setp.ne.s32 	%p56, %r14, %r104;
	@%p56 bra 	$Lt_35_163586;
	.loc	20	63	0
	mov.f32 	%f74, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_53;
$Lt_35_163586:
	.loc	20	65	0
	mov.u32 	%r105, 128;
	setp.ne.s32 	%p57, %r14, %r105;
	@%p57 bra 	$Lt_35_163842;
	.loc	20	68	0
	mov.f32 	%f74, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_53;
$Lt_35_163842:
	.loc	20	70	0
	mov.u32 	%r106, 192;
	setp.ne.s32 	%p58, %r14, %r106;
	@%p58 bra 	$Lt_35_164098;
	.loc	20	72	0
	mov.f32 	%f74, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_53;
$Lt_35_164098:
	.loc	20	76	0
	mov.f32 	%f74, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_213_53:
	.loc	20	505	0
	mov.f32 	%f77, 0f437f0000;    	// 255
	div.approx.ftz.f32 	%f78, %f77, %f74;
	mul.ftz.f32 	%f14, %f14, %f78;
	.loc	20	57	0
	@!%p3 bra 	$Lt_35_165378;
	.loc	20	59	0
	mov.f32 	%f74, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_49;
$Lt_35_165378:
	.loc	20	61	0
	mov.u32 	%r107, 64;
	setp.ne.s32 	%p59, %r14, %r107;
	@%p59 bra 	$Lt_35_165634;
	.loc	20	63	0
	mov.f32 	%f74, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_49;
$Lt_35_165634:
	.loc	20	65	0
	mov.u32 	%r108, 128;
	setp.ne.s32 	%p60, %r14, %r108;
	@%p60 bra 	$Lt_35_165890;
	.loc	20	68	0
	mov.f32 	%f74, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_49;
$Lt_35_165890:
	.loc	20	70	0
	mov.u32 	%r109, 192;
	setp.ne.s32 	%p61, %r14, %r109;
	@%p61 bra 	$Lt_35_166146;
	.loc	20	72	0
	mov.f32 	%f74, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_49;
$Lt_35_166146:
	.loc	20	76	0
	mov.f32 	%f74, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_213_49:
	.loc	20	506	0
	mov.f32 	%f79, 0f437f0000;    	// 255
	div.approx.ftz.f32 	%f80, %f79, %f74;
	mul.ftz.f32 	%f15, %f61, %f80;
	.loc	20	57	0
	@!%p3 bra 	$Lt_35_167426;
	.loc	20	59	0
	mov.f32 	%f74, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_45;
$Lt_35_167426:
	.loc	20	61	0
	mov.u32 	%r110, 64;
	setp.ne.s32 	%p62, %r14, %r110;
	@%p62 bra 	$Lt_35_167682;
	.loc	20	63	0
	mov.f32 	%f74, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_45;
$Lt_35_167682:
	.loc	20	65	0
	mov.u32 	%r111, 128;
	setp.ne.s32 	%p63, %r14, %r111;
	@%p63 bra 	$Lt_35_167938;
	.loc	20	68	0
	mov.f32 	%f74, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_45;
$Lt_35_167938:
	.loc	20	70	0
	mov.u32 	%r112, 192;
	setp.ne.s32 	%p64, %r14, %r112;
	@%p64 bra 	$Lt_35_168194;
	.loc	20	72	0
	mov.f32 	%f74, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_45;
$Lt_35_168194:
	.loc	20	76	0
	mov.f32 	%f74, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_213_45:
	.loc	20	507	0
	mov.f32 	%f81, 0f437f0000;    	// 255
	div.approx.ftz.f32 	%f82, %f81, %f74;
	mul.ftz.f32 	%f16, %f67, %f82;
	bra.uni 	$Lt_35_241666;
$Lt_35_259330:
$L_35_226818:
	.loc	20	57	0
	@!%p3 bra 	$Lt_35_168450;
	.loc	20	59	0
	mov.f32 	%f73, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_43;
$Lt_35_168450:
	.loc	20	61	0
	mov.u32 	%r113, 64;
	setp.ne.s32 	%p65, %r14, %r113;
	@%p65 bra 	$Lt_35_168706;
	.loc	20	63	0
	mov.f32 	%f73, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_43;
$Lt_35_168706:
	.loc	20	65	0
	mov.u32 	%r114, 128;
	setp.ne.s32 	%p66, %r14, %r114;
	@%p66 bra 	$Lt_35_168962;
	.loc	20	68	0
	mov.f32 	%f73, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_43;
$Lt_35_168962:
	.loc	20	70	0
	mov.u32 	%r115, 192;
	setp.ne.s32 	%p67, %r14, %r115;
	@%p67 bra 	$Lt_35_169218;
	.loc	20	72	0
	mov.f32 	%f73, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_43;
$Lt_35_169218:
	.loc	20	76	0
	mov.f32 	%f73, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_213_43:
	.loc	20	511	0
	mov.f32 	%f83, 0f437f0000;    	// 255
	div.approx.ftz.f32 	%f84, %f73, %f83;
	mul.ftz.f32 	%f12, %f84, %f12;
	mul.ftz.f32 	%f14, %f84, %f14;
	mul.ftz.f32 	%f15, %f84, %f61;
	mul.ftz.f32 	%f16, %f84, %f67;
	bra.uni 	$Lt_35_241666;
$Lt_35_84482:
	.loc	20	486	0
	setp.eq.s32 	%p68, %r21, %r14;
	@%p68 bra 	$Lt_35_241666;
	.loc	20	57	0
	mov.u32 	%r116, 0;
	setp.ne.s32 	%p69, %r21, %r116;
	@%p69 bra 	$Lt_35_170754;
	.loc	20	59	0
	mov.f32 	%f73, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_39;
$Lt_35_170754:
	.loc	20	61	0
	mov.u32 	%r117, 64;
	setp.ne.s32 	%p70, %r21, %r117;
	@%p70 bra 	$Lt_35_171010;
	.loc	20	63	0
	mov.f32 	%f73, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_39;
$Lt_35_171010:
	.loc	20	65	0
	mov.u32 	%r118, 128;
	setp.ne.s32 	%p71, %r21, %r118;
	@%p71 bra 	$Lt_35_171266;
	.loc	20	68	0
	mov.f32 	%f73, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_39;
$Lt_35_171266:
	.loc	20	70	0
	mov.u32 	%r119, 192;
	setp.ne.s32 	%p72, %r21, %r119;
	@%p72 bra 	$Lt_35_171522;
	.loc	20	72	0
	mov.f32 	%f73, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_39;
$Lt_35_171522:
	.loc	20	76	0
	mov.f32 	%f73, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_213_39:
	.loc	20	57	0
	@!%p3 bra 	$Lt_35_171778;
	.loc	20	59	0
	mov.f32 	%f74, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_37;
$Lt_35_171778:
	.loc	20	61	0
	mov.u32 	%r120, 64;
	setp.ne.s32 	%p73, %r14, %r120;
	@%p73 bra 	$Lt_35_172034;
	.loc	20	63	0
	mov.f32 	%f74, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_37;
$Lt_35_172034:
	.loc	20	65	0
	mov.u32 	%r121, 128;
	setp.ne.s32 	%p74, %r14, %r121;
	@%p74 bra 	$Lt_35_172290;
	.loc	20	68	0
	mov.f32 	%f74, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_37;
$Lt_35_172290:
	.loc	20	70	0
	mov.u32 	%r122, 192;
	setp.ne.s32 	%p75, %r14, %r122;
	@%p75 bra 	$Lt_35_172546;
	.loc	20	72	0
	mov.f32 	%f74, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_37;
$Lt_35_172546:
	.loc	20	76	0
	mov.f32 	%f74, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_213_37:
	.loc	20	517	0
	div.approx.ftz.f32 	%f85, %f73, %f74;
	mul.ftz.f32 	%f12, %f85, %f12;
	mul.ftz.f32 	%f14, %f85, %f14;
	mul.ftz.f32 	%f15, %f85, %f15;
	mul.ftz.f32 	%f16, %f85, %f16;
$Lt_35_241666:
$Lt_35_86786:
	.loc	20	520	0
	mov.u32 	%r123, 0;
	setp.eq.s32 	%p76, %r39, %r123;
	@%p76 bra 	$Lt_35_259842;
	mov.u32 	%r124, 0;
	setp.ne.s32 	%p77, %r40, %r124;
	@%p77 bra 	$Lt_35_259842;
	.loc	20	522	0
	mov.f32 	%f86, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p78, %f14, %f86;
	@!%p78 bra 	$Lt_35_242178;
	.loc	20	372	0
	neg.ftz.f32 	%f87, %f14;
	lg2.approx.ftz.f32 	%f88, %f87;
	mov.f32 	%f89, 0f400e38e4;    	// 2.22222
	mul.ftz.f32 	%f90, %f88, %f89;
	ex2.approx.ftz.f32 	%f91, %f90;
	neg.ftz.f32 	%f92, %f91;
	bra.uni 	$LDWendi___log2f_213_35;
$Lt_35_242178:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f93, %f14;
	mov.f32 	%f94, 0f400e38e4;    	// 2.22222
	mul.ftz.f32 	%f95, %f93, %f94;
	ex2.approx.ftz.f32 	%f92, %f95;
$LDWendi___log2f_213_35:
	.loc	20	522	0
	mov.f32 	%f96, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p79, %f15, %f96;
	@!%p79 bra 	$Lt_35_242690;
	.loc	20	372	0
	neg.ftz.f32 	%f97, %f15;
	lg2.approx.ftz.f32 	%f98, %f97;
	mov.f32 	%f99, 0f400e38e4;    	// 2.22222
	mul.ftz.f32 	%f100, %f98, %f99;
	ex2.approx.ftz.f32 	%f101, %f100;
	neg.ftz.f32 	%f102, %f101;
	bra.uni 	$LDWendi___log2f_213_33;
$Lt_35_242690:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f103, %f15;
	mov.f32 	%f104, 0f400e38e4;   	// 2.22222
	mul.ftz.f32 	%f105, %f103, %f104;
	ex2.approx.ftz.f32 	%f102, %f105;
$LDWendi___log2f_213_33:
	.loc	20	522	0
	mov.f32 	%f106, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p80, %f16, %f106;
	@!%p80 bra 	$Lt_35_243202;
	.loc	20	372	0
	neg.ftz.f32 	%f107, %f16;
	lg2.approx.ftz.f32 	%f108, %f107;
	mov.f32 	%f109, 0f400e38e4;   	// 2.22222
	mul.ftz.f32 	%f110, %f108, %f109;
	ex2.approx.ftz.f32 	%f111, %f110;
	neg.ftz.f32 	%f112, %f111;
	bra.uni 	$LDWendi___log2f_213_31;
$Lt_35_243202:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f113, %f16;
	mov.f32 	%f114, 0f400e38e4;   	// 2.22222
	mul.ftz.f32 	%f115, %f113, %f114;
	ex2.approx.ftz.f32 	%f112, %f115;
$LDWendi___log2f_213_31:
	.loc	20	522	0
	mov.f32 	%f14, %f92;
	mov.f32 	%f15, %f102;
	mov.f32 	%f16, %f112;
$Lt_35_259842:
$Lt_35_88834:
	.loc	20	525	0
	mov.u32 	%r125, 0;
	setp.eq.s32 	%p81, %r26, %r125;
	@%p81 bra 	$Lt_35_260354;
	mov.u32 	%r126, 0;
	setp.ne.s32 	%p82, %r31, %r126;
	@%p82 bra 	$Lt_35_260354;
	.loc	20	57	0
	mov.u32 	%r127, 0;
	setp.ne.s32 	%p83, %r21, %r127;
	@%p83 bra 	$Lt_35_173570;
	.loc	20	59	0
	mov.f32 	%f116, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_29;
$Lt_35_173570:
	.loc	20	61	0
	mov.u32 	%r128, 64;
	setp.ne.s32 	%p84, %r21, %r128;
	@%p84 bra 	$Lt_35_173826;
	.loc	20	63	0
	mov.f32 	%f116, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_29;
$Lt_35_173826:
	.loc	20	65	0
	mov.u32 	%r129, 128;
	setp.ne.s32 	%p85, %r21, %r129;
	@%p85 bra 	$Lt_35_174082;
	.loc	20	68	0
	mov.f32 	%f116, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_29;
$Lt_35_174082:
	.loc	20	70	0
	mov.u32 	%r130, 192;
	setp.ne.s32 	%p86, %r21, %r130;
	@%p86 bra 	$Lt_35_174338;
	.loc	20	72	0
	mov.f32 	%f116, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_29;
$Lt_35_174338:
	.loc	20	76	0
	mov.f32 	%f116, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_213_29:
	.loc	20	107	0
	and.b32 	%r131, %r4, 2048;
	mov.s32 	%r132, 0;
	setp.ne.s32 	%p87, %r131, %r132;
	@!%p87 bra 	$Lt_35_243970;
	.loc	20	100	0
	ld.const.f32 	%f117, [kYCbCrOffset+0];
	bra.uni 	$Lt_35_243714;
$Lt_35_243970:
	ld.const.f32 	%f117, [kYCbCrFullRangeOffset+0];
$Lt_35_243714:
	.loc	20	107	0
	@!%p87 bra 	$Lt_35_244482;
	.loc	20	100	0
	ld.const.f32 	%f118, [kYCbCrOffset+4];
	bra.uni 	$Lt_35_244226;
$Lt_35_244482:
	ld.const.f32 	%f118, [kYCbCrFullRangeOffset+4];
$Lt_35_244226:
	.loc	20	107	0
	@!%p87 bra 	$Lt_35_244994;
	.loc	20	100	0
	ld.const.f32 	%f119, [kYCbCrOffset+8];
	bra.uni 	$Lt_35_244738;
$Lt_35_244994:
	ld.const.f32 	%f119, [kYCbCrFullRangeOffset+8];
$Lt_35_244738:
	.loc	20	527	0
	mov.f32 	%f120, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f121, %f116, %f120;
	fma.rn.ftz.f32 	%f14, %f121, %f117, %f14;
	fma.rn.ftz.f32 	%f15, %f121, %f118, %f15;
	fma.rn.ftz.f32 	%f16, %f121, %f119, %f16;
$Lt_35_260354:
$Lt_35_95234:
	.loc	20	525	0
	and.b32 	%r133, %r4, 12;
	and.b32 	%r134, %r13, 12;
	setp.eq.s32 	%p88, %r133, %r134;
	@%p88 bra 	$Lt_35_245762;
	.loc	20	532	0
	mov.u32 	%r135, 8;
	setp.ne.s32 	%p89, %r134, %r135;
	@%p89 bra 	$L_35_228610;
	mov.u32 	%r136, 12;
	setp.eq.s32 	%p90, %r133, %r136;
	@%p90 bra 	$Lt_35_261122;
$L_35_228610:
	mov.u32 	%r137, 12;
	setp.eq.s32 	%p91, %r134, %r137;
	@%p91 bra 	$Lt_35_261122;
	mov.u32 	%r138, 0;
	setp.ne.s32 	%p92, %r134, %r138;
	@%p92 bra 	$L_35_227842;
$Lt_35_261122:
$L_35_228098:
	.loc	20	57	0
	mov.u32 	%r139, 0;
	setp.ne.s32 	%p93, %r21, %r139;
	@%p93 bra 	$Lt_35_176898;
	.loc	20	59	0
	mov.f32 	%f122, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_25;
$Lt_35_176898:
	.loc	20	61	0
	mov.u32 	%r140, 64;
	setp.ne.s32 	%p94, %r21, %r140;
	@%p94 bra 	$Lt_35_177154;
	.loc	20	63	0
	mov.f32 	%f122, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_25;
$Lt_35_177154:
	.loc	20	65	0
	mov.u32 	%r141, 128;
	setp.ne.s32 	%p95, %r21, %r141;
	@%p95 bra 	$Lt_35_177410;
	.loc	20	68	0
	mov.f32 	%f122, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_25;
$Lt_35_177410:
	.loc	20	70	0
	mov.u32 	%r142, 192;
	setp.ne.s32 	%p96, %r21, %r142;
	@%p96 bra 	$Lt_35_177666;
	.loc	20	72	0
	mov.f32 	%f122, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_25;
$Lt_35_177666:
	.loc	20	76	0
	mov.f32 	%f122, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_213_25:
	.loc	20	536	0
	mov.f32 	%f12, %f122;
	bra.uni 	$Lt_35_245762;
$L_35_227842:
	.loc	20	540	0
	mov.s32 	%r143, 12;
	setp.eq.s32 	%p97, %r133, %r143;
	mov.s32 	%r144, 4;
	set.eq.u32.s32 	%r145, %r134, %r144;
	neg.s32 	%r146, %r145;
	selp.s32 	%r147, 1, 0, %p97;
	mov.s32 	%r148, 8;
	set.eq.u32.s32 	%r149, %r133, %r148;
	neg.s32 	%r150, %r149;
	or.b32 	%r151, %r147, %r150;
	and.b32 	%r152, %r146, %r151;
	mov.u32 	%r153, 0;
	setp.eq.s32 	%p98, %r152, %r153;
	@%p98 bra 	$Lt_35_246018;
	.loc	20	410	0
	mov.f32 	%f123, %f14;
	mov.f32 	%f124, %f123;
	mov.f32 	%f125, %f15;
	mov.f32 	%f126, %f125;
	mov.f32 	%f127, %f16;
	mov.f32 	%f128, %f127;
	.loc	20	57	0
	mov.s32 	%r154, 0;
	setp.eq.s32 	%p40, %r21, %r154;
	@!%p40 bra 	$Lt_35_178178;
	.loc	20	59	0
	mov.f32 	%f129, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_23;
$Lt_35_178178:
	.loc	20	61	0
	mov.u32 	%r155, 64;
	setp.ne.s32 	%p99, %r21, %r155;
	@%p99 bra 	$Lt_35_178434;
	.loc	20	63	0
	mov.f32 	%f129, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_23;
$Lt_35_178434:
	.loc	20	65	0
	mov.u32 	%r156, 128;
	setp.ne.s32 	%p100, %r21, %r156;
	@%p100 bra 	$Lt_35_178690;
	.loc	20	68	0
	mov.f32 	%f129, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_23;
$Lt_35_178690:
	.loc	20	70	0
	mov.u32 	%r157, 192;
	setp.ne.s32 	%p101, %r21, %r157;
	@%p101 bra 	$Lt_35_178946;
	.loc	20	72	0
	mov.f32 	%f129, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_23;
$Lt_35_178946:
	.loc	20	76	0
	mov.f32 	%f129, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_213_23:
	.loc	20	413	0
	mov.u32 	%r158, 0;
	setp.eq.s32 	%p102, %r26, %r158;
	@%p102 bra 	$Lt_35_246274;
	.loc	20	57	0
	@!%p40 bra 	$Lt_35_179458;
	.loc	20	59	0
	mov.f32 	%f130, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_21;
$Lt_35_179458:
	.loc	20	61	0
	mov.u32 	%r159, 64;
	setp.ne.s32 	%p103, %r21, %r159;
	@%p103 bra 	$Lt_35_179714;
	.loc	20	63	0
	mov.f32 	%f130, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_21;
$Lt_35_179714:
	.loc	20	65	0
	mov.u32 	%r160, 128;
	setp.ne.s32 	%p104, %r21, %r160;
	@%p104 bra 	$Lt_35_179970;
	.loc	20	68	0
	mov.f32 	%f130, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_21;
$Lt_35_179970:
	.loc	20	70	0
	mov.u32 	%r161, 192;
	setp.ne.s32 	%p105, %r21, %r161;
	@%p105 bra 	$Lt_35_180226;
	.loc	20	72	0
	mov.f32 	%f130, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_21;
$Lt_35_180226:
	.loc	20	76	0
	mov.f32 	%f130, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_213_21:
	.loc	20	118	0
	and.b32 	%r131, %r4, 2048;
	mov.s32 	%r162, 0;
	setp.ne.s32 	%p87, %r131, %r162;
	@!%p87 bra 	$Lt_35_247042;
	.loc	20	100	0
	ld.const.f32 	%f131, [kYCbCrOffset+0];
	bra.uni 	$Lt_35_246786;
$Lt_35_247042:
	ld.const.f32 	%f131, [kYCbCrFullRangeOffset+0];
$Lt_35_246786:
	.loc	20	118	0
	@!%p87 bra 	$Lt_35_247554;
	.loc	20	100	0
	ld.const.f32 	%f132, [kYCbCrOffset+4];
	bra.uni 	$Lt_35_247298;
$Lt_35_247554:
	ld.const.f32 	%f132, [kYCbCrFullRangeOffset+4];
$Lt_35_247298:
	.loc	20	118	0
	@!%p87 bra 	$Lt_35_248066;
	.loc	20	100	0
	ld.const.f32 	%f133, [kYCbCrOffset+8];
	bra.uni 	$Lt_35_247810;
$Lt_35_248066:
	ld.const.f32 	%f133, [kYCbCrFullRangeOffset+8];
$Lt_35_247810:
	.loc	20	415	0
	mov.f32 	%f134, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f135, %f130, %f134;
	mul.ftz.f32 	%f136, %f135, %f131;
	sub.ftz.f32 	%f124, %f123, %f136;
	mul.ftz.f32 	%f137, %f135, %f132;
	sub.ftz.f32 	%f126, %f125, %f137;
	mul.ftz.f32 	%f138, %f135, %f133;
	sub.ftz.f32 	%f128, %f127, %f138;
$Lt_35_246274:
	.loc	20	418	0
	rcp.approx.ftz.f32 	%f139, %f129;
	mul.ftz.f32 	%f140, %f139, %f12;
	mul.ftz.f32 	%f124, %f140, %f124;
	.loc	20	419	0
	mul.ftz.f32 	%f126, %f140, %f126;
	.loc	20	420	0
	mul.ftz.f32 	%f128, %f140, %f128;
	.loc	20	422	0
	mov.u32 	%r163, 0;
	setp.eq.s32 	%p106, %r26, %r163;
	@%p106 bra 	$Lt_35_248322;
	.loc	20	57	0
	@!%p40 bra 	$Lt_35_182530;
	.loc	20	59	0
	mov.f32 	%f141, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_17;
$Lt_35_182530:
	.loc	20	61	0
	mov.u32 	%r164, 64;
	setp.ne.s32 	%p107, %r21, %r164;
	@%p107 bra 	$Lt_35_182786;
	.loc	20	63	0
	mov.f32 	%f141, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_17;
$Lt_35_182786:
	.loc	20	65	0
	mov.u32 	%r165, 128;
	setp.ne.s32 	%p108, %r21, %r165;
	@%p108 bra 	$Lt_35_183042;
	.loc	20	68	0
	mov.f32 	%f141, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_17;
$Lt_35_183042:
	.loc	20	70	0
	mov.u32 	%r166, 192;
	setp.ne.s32 	%p109, %r21, %r166;
	@%p109 bra 	$Lt_35_183298;
	.loc	20	72	0
	mov.f32 	%f141, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_17;
$Lt_35_183298:
	.loc	20	76	0
	mov.f32 	%f141, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_213_17:
	.loc	20	107	0
	and.b32 	%r131, %r4, 2048;
	mov.s32 	%r167, 0;
	setp.ne.s32 	%p87, %r131, %r167;
	@!%p87 bra 	$Lt_35_249090;
	.loc	20	100	0
	ld.const.f32 	%f142, [kYCbCrOffset+0];
	bra.uni 	$Lt_35_248834;
$Lt_35_249090:
	ld.const.f32 	%f142, [kYCbCrFullRangeOffset+0];
$Lt_35_248834:
	.loc	20	107	0
	@!%p87 bra 	$Lt_35_249602;
	.loc	20	100	0
	ld.const.f32 	%f143, [kYCbCrOffset+4];
	bra.uni 	$Lt_35_249346;
$Lt_35_249602:
	ld.const.f32 	%f143, [kYCbCrFullRangeOffset+4];
$Lt_35_249346:
	.loc	20	107	0
	@!%p87 bra 	$Lt_35_250114;
	.loc	20	100	0
	ld.const.f32 	%f144, [kYCbCrOffset+8];
	bra.uni 	$Lt_35_249858;
$Lt_35_250114:
	ld.const.f32 	%f144, [kYCbCrFullRangeOffset+8];
$Lt_35_249858:
	.loc	20	424	0
	mov.f32 	%f145, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f146, %f141, %f145;
	fma.rn.ftz.f32 	%f124, %f146, %f142, %f124;
	fma.rn.ftz.f32 	%f126, %f146, %f143, %f126;
	fma.rn.ftz.f32 	%f128, %f146, %f144, %f128;
$Lt_35_248322:
	.loc	20	543	0
	mov.f32 	%f14, %f124;
	mov.f32 	%f15, %f126;
	mov.f32 	%f16, %f128;
	@!%p97 bra 	$Lt_35_245762;
	.loc	20	57	0
	@!%p40 bra 	$Lt_35_185602;
	.loc	20	59	0
	mov.f32 	%f122, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_13;
$Lt_35_185602:
	.loc	20	61	0
	mov.u32 	%r168, 64;
	setp.ne.s32 	%p110, %r21, %r168;
	@%p110 bra 	$Lt_35_185858;
	.loc	20	63	0
	mov.f32 	%f122, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_13;
$Lt_35_185858:
	.loc	20	65	0
	mov.u32 	%r169, 128;
	setp.ne.s32 	%p111, %r21, %r169;
	@%p111 bra 	$Lt_35_186114;
	.loc	20	68	0
	mov.f32 	%f122, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_13;
$Lt_35_186114:
	.loc	20	70	0
	mov.u32 	%r170, 192;
	setp.ne.s32 	%p112, %r21, %r170;
	@%p112 bra 	$Lt_35_186370;
	.loc	20	72	0
	mov.f32 	%f122, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_13;
$Lt_35_186370:
	.loc	20	76	0
	mov.f32 	%f122, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_213_13:
	.loc	20	546	0
	mov.f32 	%f12, %f122;
	bra.uni 	$Lt_35_245762;
$Lt_35_246018:
	.loc	20	433	0
	mov.f32 	%f147, %f12;
	mov.f32 	%f148, %f14;
	mov.f32 	%f149, %f148;
	mov.f32 	%f150, %f15;
	mov.f32 	%f151, %f150;
	mov.f32 	%f152, %f16;
	mov.f32 	%f153, %f152;
	.loc	20	435	0
	mov.u32 	%r171, 0;
	setp.eq.s32 	%p113, %r26, %r171;
	@%p113 bra 	$Lt_35_250882;
	.loc	20	57	0
	mov.u32 	%r172, 0;
	setp.ne.s32 	%p114, %r21, %r172;
	@%p114 bra 	$Lt_35_186882;
	.loc	20	59	0
	mov.f32 	%f154, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_11;
$Lt_35_186882:
	.loc	20	61	0
	mov.u32 	%r173, 64;
	setp.ne.s32 	%p115, %r21, %r173;
	@%p115 bra 	$Lt_35_187138;
	.loc	20	63	0
	mov.f32 	%f154, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_11;
$Lt_35_187138:
	.loc	20	65	0
	mov.u32 	%r174, 128;
	setp.ne.s32 	%p116, %r21, %r174;
	@%p116 bra 	$Lt_35_187394;
	.loc	20	68	0
	mov.f32 	%f154, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_11;
$Lt_35_187394:
	.loc	20	70	0
	mov.u32 	%r175, 192;
	setp.ne.s32 	%p117, %r21, %r175;
	@%p117 bra 	$Lt_35_187650;
	.loc	20	72	0
	mov.f32 	%f154, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_11;
$Lt_35_187650:
	.loc	20	76	0
	mov.f32 	%f154, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_213_11:
	.loc	20	118	0
	and.b32 	%r131, %r4, 2048;
	mov.s32 	%r176, 0;
	setp.ne.s32 	%p87, %r131, %r176;
	@!%p87 bra 	$Lt_35_251650;
	.loc	20	100	0
	ld.const.f32 	%f155, [kYCbCrOffset+0];
	bra.uni 	$Lt_35_251394;
$Lt_35_251650:
	ld.const.f32 	%f155, [kYCbCrFullRangeOffset+0];
$Lt_35_251394:
	.loc	20	118	0
	@!%p87 bra 	$Lt_35_252162;
	.loc	20	100	0
	ld.const.f32 	%f156, [kYCbCrOffset+4];
	bra.uni 	$Lt_35_251906;
$Lt_35_252162:
	ld.const.f32 	%f156, [kYCbCrFullRangeOffset+4];
$Lt_35_251906:
	.loc	20	118	0
	@!%p87 bra 	$Lt_35_252674;
	.loc	20	100	0
	ld.const.f32 	%f157, [kYCbCrOffset+8];
	bra.uni 	$Lt_35_252418;
$Lt_35_252674:
	ld.const.f32 	%f157, [kYCbCrFullRangeOffset+8];
$Lt_35_252418:
	.loc	20	437	0
	mov.f32 	%f158, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f159, %f154, %f158;
	mul.ftz.f32 	%f160, %f159, %f155;
	sub.ftz.f32 	%f149, %f148, %f160;
	mul.ftz.f32 	%f161, %f159, %f156;
	sub.ftz.f32 	%f151, %f150, %f161;
	mul.ftz.f32 	%f162, %f159, %f157;
	sub.ftz.f32 	%f153, %f152, %f162;
$Lt_35_250882:
	mov.f32 	%f163, 0fb70637bd;   	// -8e-006
	add.ftz.f32 	%f164, %f12, %f163;
	mov.f32 	%f165, 0f00000000;   	// 0
	setp.le.ftz.f32 	%p118, %f164, %f165;
	@!%p118 bra 	$Lt_35_253186;
	mov.f32 	%f153, 0f00000000;   	// 0
	mov.f32 	%f151, 0f00000000;   	// 0
	mov.f32 	%f149, 0f00000000;   	// 0
	mov.f32 	%f147, 0f00000000;   	// 0
	bra.uni 	$Lt_35_252930;
$Lt_35_253186:
	.loc	20	57	0
	mov.u32 	%r177, 0;
	setp.ne.s32 	%p119, %r21, %r177;
	@%p119 bra 	$Lt_35_189954;
	.loc	20	59	0
	mov.f32 	%f166, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_7;
$Lt_35_189954:
	.loc	20	61	0
	mov.u32 	%r178, 64;
	setp.ne.s32 	%p120, %r21, %r178;
	@%p120 bra 	$Lt_35_190210;
	.loc	20	63	0
	mov.f32 	%f166, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_7;
$Lt_35_190210:
	.loc	20	65	0
	mov.u32 	%r179, 128;
	setp.ne.s32 	%p121, %r21, %r179;
	@%p121 bra 	$Lt_35_190466;
	.loc	20	68	0
	mov.f32 	%f166, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_7;
$Lt_35_190466:
	.loc	20	70	0
	mov.u32 	%r180, 192;
	setp.ne.s32 	%p122, %r21, %r180;
	@%p122 bra 	$Lt_35_190722;
	.loc	20	72	0
	mov.f32 	%f166, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_7;
$Lt_35_190722:
	.loc	20	76	0
	mov.f32 	%f166, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_213_7:
	.loc	20	447	0
	div.approx.ftz.f32 	%f167, %f166, %f12;
	mul.ftz.f32 	%f149, %f167, %f149;
	.loc	20	448	0
	mul.ftz.f32 	%f151, %f167, %f151;
	.loc	20	449	0
	mul.ftz.f32 	%f153, %f167, %f153;
$Lt_35_252930:
	.loc	20	452	0
	mov.u32 	%r181, 0;
	setp.eq.s32 	%p123, %r26, %r181;
	@%p123 bra 	$Lt_35_253442;
	.loc	20	57	0
	mov.u32 	%r182, 0;
	setp.ne.s32 	%p124, %r21, %r182;
	@%p124 bra 	$Lt_35_191234;
	.loc	20	59	0
	mov.f32 	%f168, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_5;
$Lt_35_191234:
	.loc	20	61	0
	mov.u32 	%r183, 64;
	setp.ne.s32 	%p125, %r21, %r183;
	@%p125 bra 	$Lt_35_191490;
	.loc	20	63	0
	mov.f32 	%f168, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_5;
$Lt_35_191490:
	.loc	20	65	0
	mov.u32 	%r184, 128;
	setp.ne.s32 	%p126, %r21, %r184;
	@%p126 bra 	$Lt_35_191746;
	.loc	20	68	0
	mov.f32 	%f168, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_5;
$Lt_35_191746:
	.loc	20	70	0
	mov.u32 	%r185, 192;
	setp.ne.s32 	%p127, %r21, %r185;
	@%p127 bra 	$Lt_35_192002;
	.loc	20	72	0
	mov.f32 	%f168, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_5;
$Lt_35_192002:
	.loc	20	76	0
	mov.f32 	%f168, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_213_5:
	.loc	20	107	0
	and.b32 	%r131, %r4, 2048;
	mov.s32 	%r186, 0;
	setp.ne.s32 	%p87, %r131, %r186;
	@!%p87 bra 	$Lt_35_254210;
	.loc	20	100	0
	ld.const.f32 	%f169, [kYCbCrOffset+0];
	bra.uni 	$Lt_35_253954;
$Lt_35_254210:
	ld.const.f32 	%f169, [kYCbCrFullRangeOffset+0];
$Lt_35_253954:
	.loc	20	107	0
	@!%p87 bra 	$Lt_35_254722;
	.loc	20	100	0
	ld.const.f32 	%f170, [kYCbCrOffset+4];
	bra.uni 	$Lt_35_254466;
$Lt_35_254722:
	ld.const.f32 	%f170, [kYCbCrFullRangeOffset+4];
$Lt_35_254466:
	.loc	20	107	0
	@!%p87 bra 	$Lt_35_255234;
	.loc	20	100	0
	ld.const.f32 	%f171, [kYCbCrOffset+8];
	bra.uni 	$Lt_35_254978;
$Lt_35_255234:
	ld.const.f32 	%f171, [kYCbCrFullRangeOffset+8];
$Lt_35_254978:
	.loc	20	454	0
	mov.f32 	%f172, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f173, %f168, %f172;
	fma.rn.ftz.f32 	%f149, %f173, %f169, %f149;
	fma.rn.ftz.f32 	%f151, %f173, %f170, %f151;
	fma.rn.ftz.f32 	%f153, %f173, %f171, %f153;
$Lt_35_253442:
	.loc	20	551	0
	mov.f32 	%f12, %f147;
	mov.f32 	%f14, %f149;
	mov.f32 	%f15, %f151;
	mov.f32 	%f16, %f153;
$Lt_35_245762:
$L_35_227586:
$Lt_35_245250:
	.loc	20	540	0
	and.b32 	%r187, %r4, 4096;
	mov.u32 	%r188, 0;
	setp.ne.s32 	%p128, %r187, %r188;
	@%p128 bra 	$Lt_35_255490;
	.loc	21	268	0
	mov.f32 	%f174, %f14;
	.loc	21	269	0
	mov.f32 	%f175, %f12;
	.loc	20	558	0
	mov.f32 	%f12, %f16;
	mov.f32 	%f14, %f15;
	mov.f32 	%f15, %f174;
	mov.f32 	%f16, %f175;
$Lt_35_255490:
	@!%p8 bra 	$Lt_35_256002;
	.loc	20	57	0
	mov.u32 	%r189, 0;
	setp.ne.s32 	%p129, %r21, %r189;
	@%p129 bra 	$Lt_35_194562;
	.loc	20	59	0
	mov.f32 	%f176, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_1;
$Lt_35_194562:
	.loc	20	61	0
	mov.u32 	%r190, 64;
	setp.ne.s32 	%p130, %r21, %r190;
	@%p130 bra 	$Lt_35_194818;
	.loc	20	63	0
	mov.f32 	%f176, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_1;
$Lt_35_194818:
	.loc	20	65	0
	mov.u32 	%r191, 128;
	setp.ne.s32 	%p131, %r21, %r191;
	@%p131 bra 	$Lt_35_195074;
	.loc	20	68	0
	mov.f32 	%f176, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_1;
$Lt_35_195074:
	.loc	20	70	0
	mov.u32 	%r192, 192;
	setp.ne.s32 	%p132, %r21, %r192;
	@%p132 bra 	$Lt_35_195330;
	.loc	20	72	0
	mov.f32 	%f176, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_213_1;
$Lt_35_195330:
	.loc	20	76	0
	mov.f32 	%f176, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_213_1:
	.loc	20	564	0
	mov.f32 	%f177, 0f3f000000;   	// 0.5
	add.ftz.f32 	%f178, %f12, %f177;
	mov.f32 	%f179, 0f00000000;   	// 0
	mov.f32 	%f180, 0f00000000;   	// 0
	setp.gt.ftz.f32 	%p133, %f178, %f180;
	selp.f32 	%f181, %f178, %f179, %p133;
	min.ftz.f32 	%f12, %f181, %f176;
	mov.f32 	%f182, 0f3f000000;   	// 0.5
	add.ftz.f32 	%f183, %f14, %f182;
	mov.f32 	%f184, 0f00000000;   	// 0
	mov.f32 	%f185, 0f00000000;   	// 0
	setp.gt.ftz.f32 	%p134, %f183, %f185;
	selp.f32 	%f186, %f183, %f184, %p134;
	min.ftz.f32 	%f14, %f186, %f176;
	mov.f32 	%f187, 0f3f000000;   	// 0.5
	add.ftz.f32 	%f188, %f15, %f187;
	mov.f32 	%f189, 0f00000000;   	// 0
	mov.f32 	%f190, 0f00000000;   	// 0
	setp.gt.ftz.f32 	%p135, %f188, %f190;
	selp.f32 	%f191, %f188, %f189, %p135;
	min.ftz.f32 	%f15, %f191, %f176;
	mov.f32 	%f192, 0f3f000000;   	// 0.5
	add.ftz.f32 	%f193, %f16, %f192;
	mov.f32 	%f194, 0f00000000;   	// 0
	mov.f32 	%f195, 0f00000000;   	// 0
	setp.gt.ftz.f32 	%p136, %f193, %f195;
	selp.f32 	%f196, %f193, %f194, %p136;
	min.ftz.f32 	%f16, %f196, %f176;
$Lt_35_256002:
	.loc	22	47	0
	mov.f32 	%f197, %f12;
	st.param.f32 	[__cudaretf__Z23ConvertPixel_422_To_4446float414IR_PixelFormatS0_+0], %f197;
	mov.f32 	%f198, %f14;
	st.param.f32 	[__cudaretf__Z23ConvertPixel_422_To_4446float414IR_PixelFormatS0_+4], %f198;
	mov.f32 	%f199, %f15;
	st.param.f32 	[__cudaretf__Z23ConvertPixel_422_To_4446float414IR_PixelFormatS0_+8], %f199;
	mov.f32 	%f200, %f16;
	st.param.f32 	[__cudaretf__Z23ConvertPixel_422_To_4446float414IR_PixelFormatS0_+12], %f200;
	ret;
$LDWend__Z23ConvertPixel_422_To_4446float414IR_PixelFormatS0_:
	} // _Z23ConvertPixel_422_To_4446float414IR_PixelFormatS0_

	.visible .func _Z4SwapIfEvRT_S1_ (.param .u64 __cudaparmf1__Z4SwapIfEvRT_S1_, .param .u64 __cudaparmf2__Z4SwapIfEvRT_S1_)
	{
	.reg .u64 %rd<6>;
	.reg .f32 %f<4>;
	.loc	23	84	0
$LDWbegin__Z4SwapIfEvRT_S1_:
	ld.param.u64 	%rd1, [__cudaparmf1__Z4SwapIfEvRT_S1_];
	mov.s64 	%rd2, %rd1;
	ld.param.u64 	%rd3, [__cudaparmf2__Z4SwapIfEvRT_S1_];
	mov.s64 	%rd4, %rd3;
	.loc	23	85	0
	ld.f32 	%f1, [%rd2+0];
	.loc	23	86	0
	ld.f32 	%f2, [%rd4+0];
	st.f32 	[%rd2+0], %f2;
	.loc	23	87	0
	st.f32 	[%rd4+0], %f1;
	.loc	23	88	0
	ret;
$LDWend__Z4SwapIfEvRT_S1_:
	} // _Z4SwapIfEvRT_S1_

	.visible .func _Z26ConvertPixelTwo_422_To_4446float4S_14IR_PixelFormatPS_S1_S0_ (.param .align 16 .b8 __cudaparmf1__Z26ConvertPixelTwo_422_To_4446float4S_14IR_PixelFormatPS_S1_S0_[16], .param .align 16 .b8 __cudaparmf2__Z26ConvertPixelTwo_422_To_4446float4S_14IR_PixelFormatPS_S1_S0_[16], .param .s32 __cudaparmf3__Z26ConvertPixelTwo_422_To_4446float4S_14IR_PixelFormatPS_S1_S0_, .param .u64 __cudaparmf4__Z26ConvertPixelTwo_422_To_4446float4S_14IR_PixelFormatPS_S1_S0_, .param .u64 __cudaparmf5__Z26ConvertPixelTwo_422_To_4446float4S_14IR_PixelFormatPS_S1_S0_, .param .s32 __cudaparmf6__Z26ConvertPixelTwo_422_To_4446float4S_14IR_PixelFormatPS_S1_S0_)
	{
	.reg .u32 %r<487>;
	.reg .u64 %rd<25>;
	.reg .f32 %f<526>;
	.reg .pred %p<370>;
	.loc	22	59	0
$LDWbegin__Z26ConvertPixelTwo_422_To_4446float4S_14IR_PixelFormatPS_S1_S0_:
	ld.param.f32 	%f1, [__cudaparmf1__Z26ConvertPixelTwo_422_To_4446float4S_14IR_PixelFormatPS_S1_S0_+0];
	mov.f32 	%f2, %f1;
	ld.param.f32 	%f3, [__cudaparmf1__Z26ConvertPixelTwo_422_To_4446float4S_14IR_PixelFormatPS_S1_S0_+4];
	mov.f32 	%f4, %f3;
	ld.param.f32 	%f5, [__cudaparmf1__Z26ConvertPixelTwo_422_To_4446float4S_14IR_PixelFormatPS_S1_S0_+8];
	mov.f32 	%f6, %f5;
	ld.param.f32 	%f7, [__cudaparmf1__Z26ConvertPixelTwo_422_To_4446float4S_14IR_PixelFormatPS_S1_S0_+12];
	mov.f32 	%f8, %f7;
	ld.param.f32 	%f9, [__cudaparmf2__Z26ConvertPixelTwo_422_To_4446float4S_14IR_PixelFormatPS_S1_S0_+0];
	mov.f32 	%f10, %f9;
	ld.param.f32 	%f11, [__cudaparmf2__Z26ConvertPixelTwo_422_To_4446float4S_14IR_PixelFormatPS_S1_S0_+4];
	mov.f32 	%f12, %f11;
	ld.param.f32 	%f13, [__cudaparmf2__Z26ConvertPixelTwo_422_To_4446float4S_14IR_PixelFormatPS_S1_S0_+8];
	mov.f32 	%f14, %f13;
	ld.param.f32 	%f15, [__cudaparmf2__Z26ConvertPixelTwo_422_To_4446float4S_14IR_PixelFormatPS_S1_S0_+12];
	mov.f32 	%f16, %f15;
	ld.param.u32 	%r1, [__cudaparmf3__Z26ConvertPixelTwo_422_To_4446float4S_14IR_PixelFormatPS_S1_S0_];
	mov.s32 	%r2, %r1;
	ld.param.u64 	%rd1, [__cudaparmf4__Z26ConvertPixelTwo_422_To_4446float4S_14IR_PixelFormatPS_S1_S0_];
	mov.s64 	%rd2, %rd1;
	ld.param.u64 	%rd3, [__cudaparmf5__Z26ConvertPixelTwo_422_To_4446float4S_14IR_PixelFormatPS_S1_S0_];
	mov.s64 	%rd4, %rd3;
	ld.param.u32 	%r3, [__cudaparmf6__Z26ConvertPixelTwo_422_To_4446float4S_14IR_PixelFormatPS_S1_S0_];
	mov.s32 	%r4, %r3;
	.loc	22	60	0
	mov.f32 	%f17, %f2;
	mov.f32 	%f18, %f4;
	mov.f32 	%f19, %f6;
	and.b32 	%r5, %r2, 4096;
	mov.s32 	%r6, 0;
	setp.ne.s32 	%p1, %r5, %r6;
	@!%p1 bra 	$Lt_37_686850;
	.loc	23	86	0
	mov.f32 	%f17, %f4;
	.loc	23	87	0
	mov.f32 	%f18, %f2;
	.loc	22	38	0
	mov.f32 	%f19, %f8;
$Lt_37_686850:
	.loc	20	57	0
	and.b32 	%r7, %r2, 1536;
	and.b32 	%r8, %r2, 448;
	mov.s32 	%r9, 512;
	setp.eq.s32 	%p2, %r7, %r9;
	mov.s32 	%r10, 2565;
	mov.s32 	%r11, 2053;
	selp.s32 	%r12, %r10, %r11, %p2;
	or.b32 	%r13, %r12, %r8;
	and.b32 	%r14, %r13, 448;
	mov.s32 	%r15, 0;
	setp.eq.s32 	%p3, %r14, %r15;
	@!%p3 bra 	$Lt_37_432898;
	.loc	20	59	0
	mov.f32 	%f20, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_233;
$Lt_37_432898:
	.loc	20	61	0
	mov.u32 	%r16, 64;
	setp.ne.s32 	%p4, %r14, %r16;
	@%p4 bra 	$Lt_37_433154;
	.loc	20	63	0
	mov.f32 	%f20, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_233;
$Lt_37_433154:
	.loc	20	65	0
	mov.u32 	%r17, 128;
	setp.ne.s32 	%p5, %r14, %r17;
	@%p5 bra 	$Lt_37_433410;
	.loc	20	68	0
	mov.f32 	%f20, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_233;
$Lt_37_433410:
	.loc	20	70	0
	mov.u32 	%r18, 192;
	setp.ne.s32 	%p6, %r14, %r18;
	@%p6 bra 	$Lt_37_433666;
	.loc	20	72	0
	mov.f32 	%f20, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_233;
$Lt_37_433666:
	.loc	20	76	0
	mov.f32 	%f20, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_233:
	.loc	20	469	0
	mov.f32 	%f21, %f19;
	mov.f32 	%f22, %f21;
	mov.f32 	%f23, %f17;
	mov.f32 	%f24, %f23;
	mov.f32 	%f25, %f18;
	mov.f32 	%f26, %f20;
	and.b32 	%r19, %r13, 4096;
	mov.s32 	%r20, 0;
	setp.eq.s32 	%p7, %r19, %r20;
	@!%p7 bra 	$Lt_37_687362;
	.loc	20	473	0
	mov.f32 	%f22, %f26;
	mov.f32 	%f24, %f25;
	mov.f32 	%f25, %f23;
	mov.f32 	%f26, %f21;
$Lt_37_687362:
	.loc	20	476	0
	and.b32 	%r21, %r4, 448;
	mov.s32 	%r22, %r4;
	mov.s32 	%r23, 256;
	setp.ne.s32 	%p8, %r21, %r23;
	and.b32 	%r24, %r22, 1;
	selp.s32 	%r25, 1, 0, %p8;
	and.b32 	%r26, %r24, %r25;
	mov.s32 	%r27, %r13;
	mov.s32 	%r28, 256;
	setp.ne.s32 	%p9, %r14, %r28;
	and.b32 	%r29, %r27, 1;
	selp.s32 	%r30, 1, 0, %p9;
	and.b32 	%r31, %r29, %r30;
	mov.u32 	%r32, 0;
	setp.eq.s32 	%p10, %r31, %r32;
	@%p10 bra 	$Lt_37_770306;
	mov.u32 	%r33, 0;
	setp.ne.s32 	%p11, %r26, %r33;
	@%p11 bra 	$Lt_37_770306;
	.loc	20	57	0
	@!%p3 bra 	$Lt_37_434178;
	.loc	20	59	0
	mov.f32 	%f27, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_231;
$Lt_37_434178:
	.loc	20	61	0
	mov.u32 	%r34, 64;
	setp.ne.s32 	%p12, %r14, %r34;
	@%p12 bra 	$Lt_37_434434;
	.loc	20	63	0
	mov.f32 	%f27, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_231;
$Lt_37_434434:
	.loc	20	65	0
	mov.u32 	%r35, 128;
	setp.ne.s32 	%p13, %r14, %r35;
	@%p13 bra 	$Lt_37_434690;
	.loc	20	68	0
	mov.f32 	%f27, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_231;
$Lt_37_434690:
	.loc	20	70	0
	mov.u32 	%r36, 192;
	setp.ne.s32 	%p14, %r14, %r36;
	@%p14 bra 	$Lt_37_434946;
	.loc	20	72	0
	mov.f32 	%f27, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_231;
$Lt_37_434946:
	.loc	20	76	0
	mov.f32 	%f27, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_231:
	.loc	20	118	0
	and.b32 	%r37, %r13, 2048;
	mov.s32 	%r38, 0;
	setp.ne.s32 	%p15, %r37, %r38;
	@!%p15 bra 	$Lt_37_688130;
	.loc	20	100	0
	ld.const.f32 	%f28, [kYCbCrOffset+0];
	bra.uni 	$Lt_37_687874;
$Lt_37_688130:
	ld.const.f32 	%f28, [kYCbCrFullRangeOffset+0];
$Lt_37_687874:
	.loc	20	118	0
	@!%p15 bra 	$Lt_37_688642;
	.loc	20	100	0
	ld.const.f32 	%f29, [kYCbCrOffset+4];
	bra.uni 	$Lt_37_688386;
$Lt_37_688642:
	ld.const.f32 	%f29, [kYCbCrFullRangeOffset+4];
$Lt_37_688386:
	.loc	20	118	0
	@!%p15 bra 	$Lt_37_689154;
	.loc	20	100	0
	ld.const.f32 	%f30, [kYCbCrOffset+8];
	bra.uni 	$Lt_37_688898;
$Lt_37_689154:
	ld.const.f32 	%f30, [kYCbCrFullRangeOffset+8];
$Lt_37_688898:
	.loc	20	478	0
	mov.f32 	%f31, 0f437f0000;    	// 255
	div.approx.ftz.f32 	%f32, %f27, %f31;
	mul.ftz.f32 	%f33, %f32, %f28;
	sub.ftz.f32 	%f24, %f24, %f33;
	mul.ftz.f32 	%f34, %f32, %f29;
	sub.ftz.f32 	%f25, %f25, %f34;
	mul.ftz.f32 	%f35, %f32, %f30;
	sub.ftz.f32 	%f26, %f26, %f35;
$Lt_37_770306:
$Lt_37_69634:
	.loc	20	481	0
	and.b32 	%r39, %r4, 2;
	and.b32 	%r40, %r13, 2;
	mov.s32 	%r41, 0;
	setp.ne.s32 	%p16, %r40, %r41;
	@!%p16 bra 	$Lt_37_770818;
	mov.u32 	%r42, 0;
	setp.ne.s32 	%p17, %r39, %r42;
	@%p17 bra 	$Lt_37_770818;
	.loc	20	483	0
	mov.f32 	%f36, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p18, %f24, %f36;
	@!%p18 bra 	$Lt_37_689410;
	.loc	20	372	0
	neg.ftz.f32 	%f37, %f24;
	lg2.approx.ftz.f32 	%f38, %f37;
	mov.f32 	%f39, 0f3ee66666;    	// 0.45
	mul.ftz.f32 	%f40, %f38, %f39;
	ex2.approx.ftz.f32 	%f41, %f40;
	neg.ftz.f32 	%f42, %f41;
	bra.uni 	$LDWendi___log2f_215_227;
$Lt_37_689410:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f43, %f24;
	mov.f32 	%f44, 0f3ee66666;    	// 0.45
	mul.ftz.f32 	%f45, %f43, %f44;
	ex2.approx.ftz.f32 	%f42, %f45;
$LDWendi___log2f_215_227:
	.loc	20	483	0
	mov.f32 	%f46, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p19, %f25, %f46;
	@!%p19 bra 	$Lt_37_689922;
	.loc	20	372	0
	neg.ftz.f32 	%f47, %f25;
	lg2.approx.ftz.f32 	%f48, %f47;
	mov.f32 	%f49, 0f3ee66666;    	// 0.45
	mul.ftz.f32 	%f50, %f48, %f49;
	ex2.approx.ftz.f32 	%f51, %f50;
	neg.ftz.f32 	%f52, %f51;
	bra.uni 	$LDWendi___log2f_215_225;
$Lt_37_689922:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f53, %f25;
	mov.f32 	%f54, 0f3ee66666;    	// 0.45
	mul.ftz.f32 	%f55, %f53, %f54;
	ex2.approx.ftz.f32 	%f52, %f55;
$LDWendi___log2f_215_225:
	.loc	20	483	0
	mov.f32 	%f56, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p20, %f26, %f56;
	@!%p20 bra 	$Lt_37_690434;
	.loc	20	372	0
	neg.ftz.f32 	%f57, %f26;
	lg2.approx.ftz.f32 	%f58, %f57;
	mov.f32 	%f59, 0f3ee66666;    	// 0.45
	mul.ftz.f32 	%f60, %f58, %f59;
	ex2.approx.ftz.f32 	%f61, %f60;
	neg.ftz.f32 	%f62, %f61;
	bra.uni 	$LDWendi___log2f_215_223;
$Lt_37_690434:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f63, %f26;
	mov.f32 	%f64, 0f3ee66666;    	// 0.45
	mul.ftz.f32 	%f65, %f63, %f64;
	ex2.approx.ftz.f32 	%f62, %f65;
$LDWendi___log2f_215_223:
	.loc	20	483	0
	mov.f32 	%f24, %f42;
	mov.f32 	%f25, %f52;
	mov.f32 	%f26, %f62;
$Lt_37_770818:
$Lt_37_73474:
	.loc	20	486	0
	and.b32 	%r43, %r4, 1;
	and.b32 	%r44, %r4, 1536;
	and.b32 	%r45, %r13, 1;
	and.b32 	%r46, %r13, 1536;
	setp.ne.s32 	%p21, %r43, %r45;
	setp.ne.s32 	%p22, %r44, %r46;
	selp.s32 	%r47, 1, 0, %p21;
	selp.s32 	%r48, 1, 0, %p22;
	or.b32 	%r49, %r47, %r48;
	mov.u32 	%r50, 0;
	setp.ne.s32 	%p23, %r49, %r50;
	@%p23 bra 	$Lt_37_124162;
	setp.eq.s32 	%p24, %r26, %r31;
	@%p24 bra 	$Lt_37_124418;
$Lt_37_124162:
	.loc	20	490	0
	mov.u32 	%r51, 0;
	setp.ne.s32 	%p25, %r45, %r51;
	@%p25 bra 	$Lt_37_691202;
	mov.s32 	%r52, 256;
	setp.eq.s32 	%p26, %r21, %r52;
	mov.u32 	%r53, 256;
	setp.ne.s32 	%p27, %r14, %r53;
	@%p27 bra 	$Lt_37_691714;
	.loc	20	137	0
	mov.s32 	%r54, 512;
	setp.eq.s32 	%p28, %r44, %r54;
	@!%p26 bra 	$Lt_37_438274;
	.loc	20	139	0
	@!%p28 bra 	$Lt_37_438530;
	.loc	20	141	0
	cvta.global.u64 	%rd5, kRGB32f_To_709YPbPr;
	mov.s64 	%rd6, %rd5;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_221;
$Lt_37_438530:
	.loc	20	145	0
	cvta.global.u64 	%rd7, kRGB32f_To_601YPbPr;
	mov.s64 	%rd6, %rd7;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_221;
$Lt_37_438274:
	.loc	20	150	0
	@!%p28 bra 	$Lt_37_438786;
	.loc	20	152	0
	cvta.global.u64 	%rd8, kRGB32f_To_709YCbCr;
	mov.s64 	%rd6, %rd8;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_221;
$Lt_37_438786:
	.loc	20	154	0
	and.b32 	%r55, %r4, 2048;
	mov.u32 	%r56, 0;
	setp.ne.s32 	%p29, %r55, %r56;
	@%p29 bra 	$Lt_37_439042;
	.loc	20	156	0
	cvta.global.u64 	%rd9, kRGB32f_To_601YCbCrFullRange;
	mov.s64 	%rd6, %rd9;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_221;
$Lt_37_439042:
	.loc	20	160	0
	cvta.global.u64 	%rd10, kRGB32f_To_601YCbCr;
	mov.s64 	%rd6, %rd10;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_221;
$Lt_37_691714:
	@!%p26 bra 	$Lt_37_692226;
	bra.uni 	$Lt_37_690946;
$Lt_37_692226:
	.loc	20	179	0
	mov.u32 	%r57, 512;
	setp.ne.s32 	%p30, %r44, %r57;
	@%p30 bra 	$Lt_37_439810;
	.loc	20	181	0
	cvta.global.u64 	%rd11, kRGB8u_To_709YCbCr;
	mov.s64 	%rd6, %rd11;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_221;
$Lt_37_439810:
	.loc	20	183	0
	and.b32 	%r58, %r4, 2048;
	mov.u32 	%r59, 0;
	setp.ne.s32 	%p31, %r58, %r59;
	@%p31 bra 	$Lt_37_440066;
	.loc	20	185	0
	cvta.global.u64 	%rd12, kRGB8u_To_601YCbCrFullRange;
	mov.s64 	%rd6, %rd12;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_221;
$Lt_37_440066:
	.loc	20	189	0
	cvta.global.u64 	%rd13, kRGB8u_To_601YCbCr;
	mov.s64 	%rd6, %rd13;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_221;
$Lt_37_691202:
	mov.s32 	%r60, 0;
	setp.eq.s32 	%p32, %r43, %r60;
	mov.u32 	%r61, 512;
	setp.ne.s32 	%p33, %r46, %r61;
	@%p33 bra 	$Lt_37_692738;
	mov.s32 	%r62, 256;
	setp.eq.s32 	%p34, %r14, %r62;
	@!%p32 bra 	$Lt_37_693250;
	mov.s32 	%r63, 256;
	setp.eq.s32 	%p26, %r21, %r63;
	@!%p34 bra 	$Lt_37_693762;
	@!%p26 bra 	$Lt_37_690946;
	.loc	20	202	0
	cvta.global.u64 	%rd14, k709YPbPr_To_RGB32f;
	mov.s64 	%rd6, %rd14;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_221;
$Lt_37_693762:
	.loc	20	211	0
	@!%p26 bra 	$Lt_37_441346;
	.loc	20	213	0
	cvta.global.u64 	%rd15, k709YCbCr_To_RGB32f;
	mov.s64 	%rd6, %rd15;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_221;
$Lt_37_441346:
	.loc	20	217	0
	cvta.global.u64 	%rd16, k709YCbCr_To_RGB8u;
	mov.s64 	%rd6, %rd16;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_221;
$Lt_37_693250:
	@!%p34 bra 	$Lt_37_694786;
	bra.uni 	$Lt_37_690946;
$Lt_37_694786:
	mov.s32 	%r64, 256;
	set.eq.u32.s32 	%r65, %r21, %r64;
	neg.s32 	%r66, %r65;
	and.b32 	%r67, %r4, 2048;
	mov.s32 	%r68, 0;
	set.eq.u32.s32 	%r69, %r67, %r68;
	neg.s32 	%r70, %r69;
	or.b32 	%r71, %r66, %r70;
	mov.u32 	%r72, 0;
	setp.eq.s32 	%p35, %r71, %r72;
	@%p35 bra 	$Lt_37_695298;
	bra.uni 	$Lt_37_690946;
$Lt_37_695298:
	cvta.global.u64 	%rd17, k709YCbCr_To_601YCbCr;
	.loc	20	250	0
	mov.s64 	%rd6, %rd17;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_221;
$Lt_37_692738:
	and.b32 	%r37, %r13, 2048;
	mov.s32 	%r73, 0;
	setp.eq.s32 	%p36, %r37, %r73;
	mov.pred 	%p37, %p36;
	mov.pred 	%p38, %p39;
	@!%p36 bra 	$Lt_37_695810;
	@!%p32 bra 	$Lt_37_690946;
	.loc	20	259	0
	mov.u32 	%r74, 256;
	setp.ne.s32 	%p40, %r21, %r74;
	@%p40 bra 	$Lt_37_442882;
	.loc	20	261	0
	cvta.global.u64 	%rd18, k601YCbCrFullRange_To_RGB32f;
	mov.s64 	%rd6, %rd18;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_221;
$Lt_37_442882:
	.loc	20	265	0
	cvta.global.u64 	%rd19, k601YCbCrFullRange_To_RGB8u;
	mov.s64 	%rd6, %rd19;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_221;
$Lt_37_695810:
	mov.s32 	%r75, 256;
	setp.eq.s32 	%p34, %r14, %r75;
	@!%p32 bra 	$Lt_37_696834;
	mov.s32 	%r76, 256;
	setp.eq.s32 	%p26, %r21, %r76;
	@!%p34 bra 	$Lt_37_697346;
	@!%p26 bra 	$Lt_37_690946;
	.loc	20	302	0
	cvta.global.u64 	%rd20, k601YPbPr_To_RGB32f;
	mov.s64 	%rd6, %rd20;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_221;
$Lt_37_697346:
	.loc	20	311	0
	@!%p26 bra 	$Lt_37_444674;
	.loc	20	313	0
	cvta.global.u64 	%rd21, k601YCbCr_To_RGB32f;
	mov.s64 	%rd6, %rd21;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_221;
$Lt_37_444674:
	.loc	20	317	0
	cvta.global.u64 	%rd22, k601YCbCr_To_RGB8u;
	mov.s64 	%rd6, %rd22;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_221;
$Lt_37_696834:
	@!%p34 bra 	$Lt_37_698370;
	bra.uni 	$Lt_37_690946;
$Lt_37_698370:
	selp.s32 	%r77, 1, 0, %p37;
	mov.s32 	%r78, 256;
	set.eq.u32.s32 	%r79, %r21, %r78;
	neg.s32 	%r80, %r79;
	or.b32 	%r81, %r77, %r80;
	mov.u32 	%r82, 0;
	setp.eq.s32 	%p41, %r81, %r82;
	@%p41 bra 	$Lt_37_698882;
	bra.uni 	$Lt_37_690946;
$Lt_37_698882:
	cvta.global.u64 	%rd23, k601YCbCr_To_709YCbCr;
	.loc	20	350	0
	mov.s64 	%rd6, %rd23;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_221;
$Lt_37_690946:
	.loc	20	355	0
	mov.u64 	%rd6, 0;
$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_221:
	.loc	20	490	0
	cvta.global.u64 	%rd7, kRGB32f_To_601YPbPr;
	cvta.global.u64 	%rd10, kRGB32f_To_601YCbCr;
	cvta.global.u64 	%rd13, kRGB8u_To_601YCbCr;
	cvta.global.u64 	%rd12, kRGB8u_To_601YCbCrFullRange;
	cvta.global.u64 	%rd9, kRGB32f_To_601YCbCrFullRange;
	cvta.global.u64 	%rd5, kRGB32f_To_709YPbPr;
	cvta.global.u64 	%rd8, kRGB32f_To_709YCbCr;
	cvta.global.u64 	%rd11, kRGB8u_To_709YCbCr;
	cvta.global.u64 	%rd17, k709YCbCr_To_601YCbCr;
	cvta.global.u64 	%rd20, k601YPbPr_To_RGB32f;
	cvta.global.u64 	%rd21, k601YCbCr_To_RGB32f;
	cvta.global.u64 	%rd22, k601YCbCr_To_RGB8u;
	cvta.global.u64 	%rd19, k601YCbCrFullRange_To_RGB8u;
	cvta.global.u64 	%rd18, k601YCbCrFullRange_To_RGB32f;
	cvta.global.u64 	%rd14, k709YPbPr_To_RGB32f;
	cvta.global.u64 	%rd15, k709YCbCr_To_RGB32f;
	cvta.global.u64 	%rd16, k709YCbCr_To_RGB8u;
	cvta.global.u64 	%rd23, k601YCbCr_To_709YCbCr;
	ld.f32 	%f66, [%rd6+16];
	mul.ftz.f32 	%f67, %f66, %f25;
	ld.f32 	%f68, [%rd6+12];
	fma.rn.ftz.f32 	%f69, %f68, %f24, %f67;
	ld.f32 	%f70, [%rd6+20];
	fma.rn.ftz.f32 	%f71, %f70, %f26, %f69;
	ld.f32 	%f72, [%rd6+28];
	mul.ftz.f32 	%f73, %f72, %f25;
	ld.f32 	%f74, [%rd6+24];
	fma.rn.ftz.f32 	%f75, %f74, %f24, %f73;
	ld.f32 	%f76, [%rd6+32];
	fma.rn.ftz.f32 	%f77, %f76, %f26, %f75;
	ld.f32 	%f78, [%rd6+4];
	mul.ftz.f32 	%f79, %f78, %f25;
	ld.f32 	%f80, [%rd6+0];
	fma.rn.ftz.f32 	%f81, %f80, %f24, %f79;
	ld.f32 	%f82, [%rd6+8];
	fma.rn.ftz.f32 	%f24, %f82, %f26, %f81;
	mov.f32 	%f25, %f71;
	mov.f32 	%f26, %f77;
	setp.ne.s32 	%p42, %r21, %r14;
	@!%p42 bra 	$Lt_37_126722;
	.loc	20	494	0
	mov.s32 	%r83, 256;
	setp.eq.s32 	%p34, %r14, %r83;
	@!%p34 bra 	$L_37_677634;
	mov.s32 	%r84, 0;
	setp.eq.s32 	%p43, %r21, %r84;
	@%p43 bra 	$L_37_677378;
$L_37_677634:
	@!%p3 bra 	$Lt_37_772098;
	mov.u32 	%r85, 256;
	setp.ne.s32 	%p44, %r21, %r85;
	@%p44 bra 	$Lt_37_772098;
	mov.s32 	%r86, 0;
	setp.eq.s32 	%p43, %r21, %r86;
$L_37_677378:
	.loc	20	57	0
	@!%p43 bra 	$Lt_37_446210;
	.loc	20	59	0
	mov.f32 	%f83, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_219;
$Lt_37_446210:
	.loc	20	61	0
	mov.u32 	%r87, 64;
	setp.ne.s32 	%p45, %r21, %r87;
	@%p45 bra 	$Lt_37_446466;
	.loc	20	63	0
	mov.f32 	%f83, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_219;
$Lt_37_446466:
	.loc	20	65	0
	mov.u32 	%r88, 128;
	setp.ne.s32 	%p46, %r21, %r88;
	@%p46 bra 	$Lt_37_446722;
	.loc	20	68	0
	mov.f32 	%f83, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_219;
$Lt_37_446722:
	.loc	20	70	0
	mov.u32 	%r89, 192;
	setp.ne.s32 	%p47, %r21, %r89;
	@%p47 bra 	$Lt_37_446978;
	.loc	20	72	0
	mov.f32 	%f83, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_219;
$Lt_37_446978:
	.loc	20	76	0
	mov.f32 	%f83, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_219:
	.loc	20	57	0
	@!%p3 bra 	$Lt_37_447234;
	.loc	20	59	0
	mov.f32 	%f84, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_217;
$Lt_37_447234:
	.loc	20	61	0
	mov.u32 	%r90, 64;
	setp.ne.s32 	%p48, %r14, %r90;
	@%p48 bra 	$Lt_37_447490;
	.loc	20	63	0
	mov.f32 	%f84, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_217;
$Lt_37_447490:
	.loc	20	65	0
	mov.u32 	%r91, 128;
	setp.ne.s32 	%p49, %r14, %r91;
	@%p49 bra 	$Lt_37_447746;
	.loc	20	68	0
	mov.f32 	%f84, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_217;
$Lt_37_447746:
	.loc	20	70	0
	mov.u32 	%r92, 192;
	setp.ne.s32 	%p50, %r14, %r92;
	@%p50 bra 	$Lt_37_448002;
	.loc	20	72	0
	mov.f32 	%f84, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_217;
$Lt_37_448002:
	.loc	20	76	0
	mov.f32 	%f84, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_217:
	.loc	20	498	0
	div.approx.ftz.f32 	%f85, %f83, %f84;
	mul.ftz.f32 	%f22, %f22, %f85;
	bra.uni 	$Lt_37_126722;
$Lt_37_772098:
$L_37_677122:
	.loc	20	500	0
	@!%p34 bra 	$L_37_678658;
	@%p8 bra 	$L_37_678402;
$L_37_678658:
	@!%p9 bra 	$Lt_37_773122;
	mov.u32 	%r93, 256;
	setp.ne.s32 	%p51, %r21, %r93;
	@%p51 bra 	$Lt_37_773122;
$L_37_678402:
	.loc	20	57	0
	mov.u32 	%r94, 0;
	setp.ne.s32 	%p52, %r21, %r94;
	@%p52 bra 	$Lt_37_448514;
	.loc	20	59	0
	mov.f32 	%f83, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_215;
$Lt_37_448514:
	.loc	20	61	0
	mov.u32 	%r95, 64;
	setp.ne.s32 	%p53, %r21, %r95;
	@%p53 bra 	$Lt_37_448770;
	.loc	20	63	0
	mov.f32 	%f83, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_215;
$Lt_37_448770:
	.loc	20	65	0
	mov.u32 	%r96, 128;
	setp.ne.s32 	%p54, %r21, %r96;
	@%p54 bra 	$Lt_37_449026;
	.loc	20	68	0
	mov.f32 	%f83, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_215;
$Lt_37_449026:
	.loc	20	70	0
	mov.u32 	%r97, 192;
	setp.ne.s32 	%p55, %r21, %r97;
	@%p55 bra 	$Lt_37_449282;
	.loc	20	72	0
	mov.f32 	%f83, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_215;
$Lt_37_449282:
	.loc	20	76	0
	mov.f32 	%f83, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_215:
	.loc	20	57	0
	@!%p3 bra 	$Lt_37_449538;
	.loc	20	59	0
	mov.f32 	%f84, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_213;
$Lt_37_449538:
	.loc	20	61	0
	mov.u32 	%r98, 64;
	setp.ne.s32 	%p56, %r14, %r98;
	@%p56 bra 	$Lt_37_449794;
	.loc	20	63	0
	mov.f32 	%f84, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_213;
$Lt_37_449794:
	.loc	20	65	0
	mov.u32 	%r99, 128;
	setp.ne.s32 	%p57, %r14, %r99;
	@%p57 bra 	$Lt_37_450050;
	.loc	20	68	0
	mov.f32 	%f84, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_213;
$Lt_37_450050:
	.loc	20	70	0
	mov.u32 	%r100, 192;
	setp.ne.s32 	%p58, %r14, %r100;
	@%p58 bra 	$Lt_37_450306;
	.loc	20	72	0
	mov.f32 	%f84, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_213;
$Lt_37_450306:
	.loc	20	76	0
	mov.f32 	%f84, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_213:
	.loc	20	504	0
	div.approx.ftz.f32 	%f86, %f83, %f84;
	mul.ftz.f32 	%f22, %f22, %f86;
	.loc	20	57	0
	@!%p3 bra 	$Lt_37_451586;
	.loc	20	59	0
	mov.f32 	%f84, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_209;
$Lt_37_451586:
	.loc	20	61	0
	mov.u32 	%r101, 64;
	setp.ne.s32 	%p59, %r14, %r101;
	@%p59 bra 	$Lt_37_451842;
	.loc	20	63	0
	mov.f32 	%f84, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_209;
$Lt_37_451842:
	.loc	20	65	0
	mov.u32 	%r102, 128;
	setp.ne.s32 	%p60, %r14, %r102;
	@%p60 bra 	$Lt_37_452098;
	.loc	20	68	0
	mov.f32 	%f84, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_209;
$Lt_37_452098:
	.loc	20	70	0
	mov.u32 	%r103, 192;
	setp.ne.s32 	%p61, %r14, %r103;
	@%p61 bra 	$Lt_37_452354;
	.loc	20	72	0
	mov.f32 	%f84, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_209;
$Lt_37_452354:
	.loc	20	76	0
	mov.f32 	%f84, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_209:
	.loc	20	505	0
	mov.f32 	%f87, 0f437f0000;    	// 255
	div.approx.ftz.f32 	%f88, %f87, %f84;
	mul.ftz.f32 	%f24, %f24, %f88;
	.loc	20	57	0
	@!%p3 bra 	$Lt_37_453634;
	.loc	20	59	0
	mov.f32 	%f84, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_205;
$Lt_37_453634:
	.loc	20	61	0
	mov.u32 	%r104, 64;
	setp.ne.s32 	%p62, %r14, %r104;
	@%p62 bra 	$Lt_37_453890;
	.loc	20	63	0
	mov.f32 	%f84, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_205;
$Lt_37_453890:
	.loc	20	65	0
	mov.u32 	%r105, 128;
	setp.ne.s32 	%p63, %r14, %r105;
	@%p63 bra 	$Lt_37_454146;
	.loc	20	68	0
	mov.f32 	%f84, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_205;
$Lt_37_454146:
	.loc	20	70	0
	mov.u32 	%r106, 192;
	setp.ne.s32 	%p64, %r14, %r106;
	@%p64 bra 	$Lt_37_454402;
	.loc	20	72	0
	mov.f32 	%f84, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_205;
$Lt_37_454402:
	.loc	20	76	0
	mov.f32 	%f84, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_205:
	.loc	20	506	0
	mov.f32 	%f89, 0f437f0000;    	// 255
	div.approx.ftz.f32 	%f90, %f89, %f84;
	mul.ftz.f32 	%f25, %f71, %f90;
	.loc	20	57	0
	@!%p3 bra 	$Lt_37_455682;
	.loc	20	59	0
	mov.f32 	%f84, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_201;
$Lt_37_455682:
	.loc	20	61	0
	mov.u32 	%r107, 64;
	setp.ne.s32 	%p65, %r14, %r107;
	@%p65 bra 	$Lt_37_455938;
	.loc	20	63	0
	mov.f32 	%f84, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_201;
$Lt_37_455938:
	.loc	20	65	0
	mov.u32 	%r108, 128;
	setp.ne.s32 	%p66, %r14, %r108;
	@%p66 bra 	$Lt_37_456194;
	.loc	20	68	0
	mov.f32 	%f84, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_201;
$Lt_37_456194:
	.loc	20	70	0
	mov.u32 	%r109, 192;
	setp.ne.s32 	%p67, %r14, %r109;
	@%p67 bra 	$Lt_37_456450;
	.loc	20	72	0
	mov.f32 	%f84, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_201;
$Lt_37_456450:
	.loc	20	76	0
	mov.f32 	%f84, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_201:
	.loc	20	507	0
	mov.f32 	%f91, 0f437f0000;    	// 255
	div.approx.ftz.f32 	%f92, %f91, %f84;
	mul.ftz.f32 	%f26, %f77, %f92;
	bra.uni 	$Lt_37_126722;
$Lt_37_773122:
$L_37_678146:
	.loc	20	57	0
	@!%p3 bra 	$Lt_37_456706;
	.loc	20	59	0
	mov.f32 	%f83, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_199;
$Lt_37_456706:
	.loc	20	61	0
	mov.u32 	%r110, 64;
	setp.ne.s32 	%p68, %r14, %r110;
	@%p68 bra 	$Lt_37_456962;
	.loc	20	63	0
	mov.f32 	%f83, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_199;
$Lt_37_456962:
	.loc	20	65	0
	mov.u32 	%r111, 128;
	setp.ne.s32 	%p69, %r14, %r111;
	@%p69 bra 	$Lt_37_457218;
	.loc	20	68	0
	mov.f32 	%f83, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_199;
$Lt_37_457218:
	.loc	20	70	0
	mov.u32 	%r112, 192;
	setp.ne.s32 	%p70, %r14, %r112;
	@%p70 bra 	$Lt_37_457474;
	.loc	20	72	0
	mov.f32 	%f83, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_199;
$Lt_37_457474:
	.loc	20	76	0
	mov.f32 	%f83, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_199:
	.loc	20	511	0
	mov.f32 	%f93, 0f437f0000;    	// 255
	div.approx.ftz.f32 	%f94, %f83, %f93;
	mul.ftz.f32 	%f22, %f94, %f22;
	mul.ftz.f32 	%f24, %f94, %f24;
	mul.ftz.f32 	%f25, %f94, %f71;
	mul.ftz.f32 	%f26, %f94, %f77;
	bra.uni 	$Lt_37_126722;
$Lt_37_124418:
	.loc	20	486	0
	setp.ne.s32 	%p42, %r21, %r14;
	@!%p42 bra 	$Lt_37_699650;
	.loc	20	57	0
	mov.u32 	%r113, 0;
	setp.ne.s32 	%p71, %r21, %r113;
	@%p71 bra 	$Lt_37_459010;
	.loc	20	59	0
	mov.f32 	%f83, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_195;
$Lt_37_459010:
	.loc	20	61	0
	mov.u32 	%r114, 64;
	setp.ne.s32 	%p72, %r21, %r114;
	@%p72 bra 	$Lt_37_459266;
	.loc	20	63	0
	mov.f32 	%f83, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_195;
$Lt_37_459266:
	.loc	20	65	0
	mov.u32 	%r115, 128;
	setp.ne.s32 	%p73, %r21, %r115;
	@%p73 bra 	$Lt_37_459522;
	.loc	20	68	0
	mov.f32 	%f83, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_195;
$Lt_37_459522:
	.loc	20	70	0
	mov.u32 	%r116, 192;
	setp.ne.s32 	%p74, %r21, %r116;
	@%p74 bra 	$Lt_37_459778;
	.loc	20	72	0
	mov.f32 	%f83, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_195;
$Lt_37_459778:
	.loc	20	76	0
	mov.f32 	%f83, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_195:
	.loc	20	57	0
	@!%p3 bra 	$Lt_37_460034;
	.loc	20	59	0
	mov.f32 	%f84, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_193;
$Lt_37_460034:
	.loc	20	61	0
	mov.u32 	%r117, 64;
	setp.ne.s32 	%p75, %r14, %r117;
	@%p75 bra 	$Lt_37_460290;
	.loc	20	63	0
	mov.f32 	%f84, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_193;
$Lt_37_460290:
	.loc	20	65	0
	mov.u32 	%r118, 128;
	setp.ne.s32 	%p76, %r14, %r118;
	@%p76 bra 	$Lt_37_460546;
	.loc	20	68	0
	mov.f32 	%f84, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_193;
$Lt_37_460546:
	.loc	20	70	0
	mov.u32 	%r119, 192;
	setp.ne.s32 	%p77, %r14, %r119;
	@%p77 bra 	$Lt_37_460802;
	.loc	20	72	0
	mov.f32 	%f84, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_193;
$Lt_37_460802:
	.loc	20	76	0
	mov.f32 	%f84, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_193:
	.loc	20	517	0
	div.approx.ftz.f32 	%f95, %f83, %f84;
	mul.ftz.f32 	%f22, %f95, %f22;
	mul.ftz.f32 	%f24, %f95, %f24;
	mul.ftz.f32 	%f25, %f95, %f25;
	mul.ftz.f32 	%f26, %f95, %f26;
$Lt_37_699650:
	cvta.global.u64 	%rd23, k601YCbCr_To_709YCbCr;
	cvta.global.u64 	%rd16, k709YCbCr_To_RGB8u;
	cvta.global.u64 	%rd15, k709YCbCr_To_RGB32f;
	cvta.global.u64 	%rd14, k709YPbPr_To_RGB32f;
	cvta.global.u64 	%rd18, k601YCbCrFullRange_To_RGB32f;
	cvta.global.u64 	%rd19, k601YCbCrFullRange_To_RGB8u;
	cvta.global.u64 	%rd22, k601YCbCr_To_RGB8u;
	cvta.global.u64 	%rd21, k601YCbCr_To_RGB32f;
	cvta.global.u64 	%rd20, k601YPbPr_To_RGB32f;
	cvta.global.u64 	%rd17, k709YCbCr_To_601YCbCr;
	cvta.global.u64 	%rd11, kRGB8u_To_709YCbCr;
	cvta.global.u64 	%rd8, kRGB32f_To_709YCbCr;
	cvta.global.u64 	%rd5, kRGB32f_To_709YPbPr;
	cvta.global.u64 	%rd9, kRGB32f_To_601YCbCrFullRange;
	cvta.global.u64 	%rd12, kRGB8u_To_601YCbCrFullRange;
	cvta.global.u64 	%rd13, kRGB8u_To_601YCbCr;
	cvta.global.u64 	%rd10, kRGB32f_To_601YCbCr;
	cvta.global.u64 	%rd7, kRGB32f_To_601YPbPr;
$Lt_37_126722:
	.loc	20	520	0
	mov.s32 	%r120, 0;
	setp.ne.s32 	%p78, %r39, %r120;
	@!%p78 bra 	$Lt_37_773634;
	mov.u32 	%r121, 0;
	setp.ne.s32 	%p79, %r40, %r121;
	@%p79 bra 	$Lt_37_773634;
	.loc	20	522	0
	mov.f32 	%f96, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p80, %f24, %f96;
	@!%p80 bra 	$Lt_37_700162;
	.loc	20	372	0
	neg.ftz.f32 	%f97, %f24;
	lg2.approx.ftz.f32 	%f98, %f97;
	mov.f32 	%f99, 0f400e38e4;    	// 2.22222
	mul.ftz.f32 	%f100, %f98, %f99;
	ex2.approx.ftz.f32 	%f101, %f100;
	neg.ftz.f32 	%f102, %f101;
	bra.uni 	$LDWendi___log2f_215_191;
$Lt_37_700162:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f103, %f24;
	mov.f32 	%f104, 0f400e38e4;   	// 2.22222
	mul.ftz.f32 	%f105, %f103, %f104;
	ex2.approx.ftz.f32 	%f102, %f105;
$LDWendi___log2f_215_191:
	.loc	20	522	0
	mov.f32 	%f106, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p81, %f25, %f106;
	@!%p81 bra 	$Lt_37_700674;
	.loc	20	372	0
	neg.ftz.f32 	%f107, %f25;
	lg2.approx.ftz.f32 	%f108, %f107;
	mov.f32 	%f109, 0f400e38e4;   	// 2.22222
	mul.ftz.f32 	%f110, %f108, %f109;
	ex2.approx.ftz.f32 	%f111, %f110;
	neg.ftz.f32 	%f112, %f111;
	bra.uni 	$LDWendi___log2f_215_189;
$Lt_37_700674:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f113, %f25;
	mov.f32 	%f114, 0f400e38e4;   	// 2.22222
	mul.ftz.f32 	%f115, %f113, %f114;
	ex2.approx.ftz.f32 	%f112, %f115;
$LDWendi___log2f_215_189:
	.loc	20	522	0
	mov.f32 	%f116, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p82, %f26, %f116;
	@!%p82 bra 	$Lt_37_701186;
	.loc	20	372	0
	neg.ftz.f32 	%f117, %f26;
	lg2.approx.ftz.f32 	%f118, %f117;
	mov.f32 	%f119, 0f400e38e4;   	// 2.22222
	mul.ftz.f32 	%f120, %f118, %f119;
	ex2.approx.ftz.f32 	%f121, %f120;
	neg.ftz.f32 	%f122, %f121;
	bra.uni 	$LDWendi___log2f_215_187;
$Lt_37_701186:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f123, %f26;
	mov.f32 	%f124, 0f400e38e4;   	// 2.22222
	mul.ftz.f32 	%f125, %f123, %f124;
	ex2.approx.ftz.f32 	%f122, %f125;
$LDWendi___log2f_215_187:
	.loc	20	522	0
	mov.f32 	%f24, %f102;
	mov.f32 	%f25, %f112;
	mov.f32 	%f26, %f122;
$Lt_37_773634:
$Lt_37_128770:
	.loc	20	525	0
	mov.u32 	%r122, 0;
	setp.eq.s32 	%p83, %r26, %r122;
	@%p83 bra 	$Lt_37_774146;
	mov.u32 	%r123, 0;
	setp.ne.s32 	%p84, %r31, %r123;
	@%p84 bra 	$Lt_37_774146;
	.loc	20	57	0
	mov.u32 	%r124, 0;
	setp.ne.s32 	%p85, %r21, %r124;
	@%p85 bra 	$Lt_37_461826;
	.loc	20	59	0
	mov.f32 	%f126, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_185;
$Lt_37_461826:
	.loc	20	61	0
	mov.u32 	%r125, 64;
	setp.ne.s32 	%p86, %r21, %r125;
	@%p86 bra 	$Lt_37_462082;
	.loc	20	63	0
	mov.f32 	%f126, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_185;
$Lt_37_462082:
	.loc	20	65	0
	mov.u32 	%r126, 128;
	setp.ne.s32 	%p87, %r21, %r126;
	@%p87 bra 	$Lt_37_462338;
	.loc	20	68	0
	mov.f32 	%f126, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_185;
$Lt_37_462338:
	.loc	20	70	0
	mov.u32 	%r127, 192;
	setp.ne.s32 	%p88, %r21, %r127;
	@%p88 bra 	$Lt_37_462594;
	.loc	20	72	0
	mov.f32 	%f126, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_185;
$Lt_37_462594:
	.loc	20	76	0
	mov.f32 	%f126, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_185:
	.loc	20	107	0
	and.b32 	%r128, %r4, 2048;
	mov.s32 	%r129, 0;
	setp.ne.s32 	%p89, %r128, %r129;
	@!%p89 bra 	$Lt_37_701954;
	.loc	20	100	0
	ld.const.f32 	%f127, [kYCbCrOffset+0];
	bra.uni 	$Lt_37_701698;
$Lt_37_701954:
	ld.const.f32 	%f127, [kYCbCrFullRangeOffset+0];
$Lt_37_701698:
	.loc	20	107	0
	@!%p89 bra 	$Lt_37_702466;
	.loc	20	100	0
	ld.const.f32 	%f128, [kYCbCrOffset+4];
	bra.uni 	$Lt_37_702210;
$Lt_37_702466:
	ld.const.f32 	%f128, [kYCbCrFullRangeOffset+4];
$Lt_37_702210:
	.loc	20	107	0
	@!%p89 bra 	$Lt_37_702978;
	.loc	20	100	0
	ld.const.f32 	%f129, [kYCbCrOffset+8];
	bra.uni 	$Lt_37_702722;
$Lt_37_702978:
	ld.const.f32 	%f129, [kYCbCrFullRangeOffset+8];
$Lt_37_702722:
	.loc	20	527	0
	mov.f32 	%f130, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f131, %f126, %f130;
	fma.rn.ftz.f32 	%f24, %f131, %f127, %f24;
	fma.rn.ftz.f32 	%f25, %f131, %f128, %f25;
	fma.rn.ftz.f32 	%f26, %f131, %f129, %f26;
$Lt_37_774146:
$Lt_37_135170:
	.loc	20	525	0
	and.b32 	%r130, %r4, 12;
	and.b32 	%r131, %r13, 12;
	setp.ne.s32 	%p90, %r130, %r131;
	@!%p90 bra 	$Lt_37_703746;
	.loc	20	532	0
	mov.u32 	%r132, 8;
	setp.ne.s32 	%p91, %r131, %r132;
	@%p91 bra 	$L_37_679938;
	mov.u32 	%r133, 12;
	setp.eq.s32 	%p92, %r130, %r133;
	@%p92 bra 	$Lt_37_774914;
$L_37_679938:
	mov.u32 	%r134, 12;
	setp.eq.s32 	%p93, %r131, %r134;
	@%p93 bra 	$Lt_37_774914;
	mov.u32 	%r135, 0;
	setp.ne.s32 	%p94, %r131, %r135;
	@%p94 bra 	$L_37_679170;
$Lt_37_774914:
$L_37_679426:
	.loc	20	57	0
	mov.u32 	%r136, 0;
	setp.ne.s32 	%p95, %r21, %r136;
	@%p95 bra 	$Lt_37_465154;
	.loc	20	59	0
	mov.f32 	%f132, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_181;
$Lt_37_465154:
	.loc	20	61	0
	mov.u32 	%r137, 64;
	setp.ne.s32 	%p96, %r21, %r137;
	@%p96 bra 	$Lt_37_465410;
	.loc	20	63	0
	mov.f32 	%f132, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_181;
$Lt_37_465410:
	.loc	20	65	0
	mov.u32 	%r138, 128;
	setp.ne.s32 	%p97, %r21, %r138;
	@%p97 bra 	$Lt_37_465666;
	.loc	20	68	0
	mov.f32 	%f132, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_181;
$Lt_37_465666:
	.loc	20	70	0
	mov.u32 	%r139, 192;
	setp.ne.s32 	%p98, %r21, %r139;
	@%p98 bra 	$Lt_37_465922;
	.loc	20	72	0
	mov.f32 	%f132, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_181;
$Lt_37_465922:
	.loc	20	76	0
	mov.f32 	%f132, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_181:
	.loc	20	536	0
	mov.f32 	%f22, %f132;
	bra.uni 	$Lt_37_703746;
$L_37_679170:
	.loc	20	540	0
	mov.s32 	%r140, 12;
	setp.eq.s32 	%p99, %r130, %r140;
	mov.s32 	%r141, 4;
	set.eq.u32.s32 	%r142, %r131, %r141;
	neg.s32 	%r143, %r142;
	selp.s32 	%r144, 1, 0, %p99;
	mov.s32 	%r145, 8;
	set.eq.u32.s32 	%r146, %r130, %r145;
	neg.s32 	%r147, %r146;
	or.b32 	%r148, %r144, %r147;
	and.b32 	%r149, %r143, %r148;
	mov.u32 	%r150, 0;
	setp.eq.s32 	%p100, %r149, %r150;
	@%p100 bra 	$Lt_37_704002;
	.loc	20	410	0
	mov.f32 	%f133, %f24;
	mov.f32 	%f134, %f133;
	mov.f32 	%f135, %f25;
	mov.f32 	%f136, %f135;
	mov.f32 	%f137, %f26;
	mov.f32 	%f138, %f137;
	.loc	20	57	0
	mov.s32 	%r151, 0;
	setp.eq.s32 	%p43, %r21, %r151;
	@!%p43 bra 	$Lt_37_466434;
	.loc	20	59	0
	mov.f32 	%f139, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_179;
$Lt_37_466434:
	.loc	20	61	0
	mov.u32 	%r152, 64;
	setp.ne.s32 	%p101, %r21, %r152;
	@%p101 bra 	$Lt_37_466690;
	.loc	20	63	0
	mov.f32 	%f139, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_179;
$Lt_37_466690:
	.loc	20	65	0
	mov.u32 	%r153, 128;
	setp.ne.s32 	%p102, %r21, %r153;
	@%p102 bra 	$Lt_37_466946;
	.loc	20	68	0
	mov.f32 	%f139, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_179;
$Lt_37_466946:
	.loc	20	70	0
	mov.u32 	%r154, 192;
	setp.ne.s32 	%p103, %r21, %r154;
	@%p103 bra 	$Lt_37_467202;
	.loc	20	72	0
	mov.f32 	%f139, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_179;
$Lt_37_467202:
	.loc	20	76	0
	mov.f32 	%f139, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_179:
	.loc	20	413	0
	mov.u32 	%r155, 0;
	setp.eq.s32 	%p104, %r26, %r155;
	@%p104 bra 	$Lt_37_704258;
	.loc	20	57	0
	@!%p43 bra 	$Lt_37_467714;
	.loc	20	59	0
	mov.f32 	%f140, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_177;
$Lt_37_467714:
	.loc	20	61	0
	mov.u32 	%r156, 64;
	setp.ne.s32 	%p105, %r21, %r156;
	@%p105 bra 	$Lt_37_467970;
	.loc	20	63	0
	mov.f32 	%f140, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_177;
$Lt_37_467970:
	.loc	20	65	0
	mov.u32 	%r157, 128;
	setp.ne.s32 	%p106, %r21, %r157;
	@%p106 bra 	$Lt_37_468226;
	.loc	20	68	0
	mov.f32 	%f140, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_177;
$Lt_37_468226:
	.loc	20	70	0
	mov.u32 	%r158, 192;
	setp.ne.s32 	%p107, %r21, %r158;
	@%p107 bra 	$Lt_37_468482;
	.loc	20	72	0
	mov.f32 	%f140, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_177;
$Lt_37_468482:
	.loc	20	76	0
	mov.f32 	%f140, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_177:
	.loc	20	118	0
	and.b32 	%r128, %r4, 2048;
	mov.s32 	%r159, 0;
	setp.ne.s32 	%p89, %r128, %r159;
	@!%p89 bra 	$Lt_37_705026;
	.loc	20	100	0
	ld.const.f32 	%f141, [kYCbCrOffset+0];
	bra.uni 	$Lt_37_704770;
$Lt_37_705026:
	ld.const.f32 	%f141, [kYCbCrFullRangeOffset+0];
$Lt_37_704770:
	.loc	20	118	0
	@!%p89 bra 	$Lt_37_705538;
	.loc	20	100	0
	ld.const.f32 	%f142, [kYCbCrOffset+4];
	bra.uni 	$Lt_37_705282;
$Lt_37_705538:
	ld.const.f32 	%f142, [kYCbCrFullRangeOffset+4];
$Lt_37_705282:
	.loc	20	118	0
	@!%p89 bra 	$Lt_37_706050;
	.loc	20	100	0
	ld.const.f32 	%f143, [kYCbCrOffset+8];
	bra.uni 	$Lt_37_705794;
$Lt_37_706050:
	ld.const.f32 	%f143, [kYCbCrFullRangeOffset+8];
$Lt_37_705794:
	.loc	20	415	0
	mov.f32 	%f144, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f145, %f140, %f144;
	mul.ftz.f32 	%f146, %f145, %f141;
	sub.ftz.f32 	%f134, %f133, %f146;
	mul.ftz.f32 	%f147, %f145, %f142;
	sub.ftz.f32 	%f136, %f135, %f147;
	mul.ftz.f32 	%f148, %f145, %f143;
	sub.ftz.f32 	%f138, %f137, %f148;
$Lt_37_704258:
	.loc	20	418	0
	rcp.approx.ftz.f32 	%f149, %f139;
	mul.ftz.f32 	%f150, %f149, %f22;
	mul.ftz.f32 	%f134, %f150, %f134;
	.loc	20	419	0
	mul.ftz.f32 	%f136, %f150, %f136;
	.loc	20	420	0
	mul.ftz.f32 	%f138, %f150, %f138;
	.loc	20	422	0
	mov.u32 	%r160, 0;
	setp.eq.s32 	%p108, %r26, %r160;
	@%p108 bra 	$Lt_37_706306;
	.loc	20	57	0
	@!%p43 bra 	$Lt_37_470786;
	.loc	20	59	0
	mov.f32 	%f151, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_173;
$Lt_37_470786:
	.loc	20	61	0
	mov.u32 	%r161, 64;
	setp.ne.s32 	%p109, %r21, %r161;
	@%p109 bra 	$Lt_37_471042;
	.loc	20	63	0
	mov.f32 	%f151, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_173;
$Lt_37_471042:
	.loc	20	65	0
	mov.u32 	%r162, 128;
	setp.ne.s32 	%p110, %r21, %r162;
	@%p110 bra 	$Lt_37_471298;
	.loc	20	68	0
	mov.f32 	%f151, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_173;
$Lt_37_471298:
	.loc	20	70	0
	mov.u32 	%r163, 192;
	setp.ne.s32 	%p111, %r21, %r163;
	@%p111 bra 	$Lt_37_471554;
	.loc	20	72	0
	mov.f32 	%f151, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_173;
$Lt_37_471554:
	.loc	20	76	0
	mov.f32 	%f151, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_173:
	.loc	20	107	0
	and.b32 	%r128, %r4, 2048;
	mov.s32 	%r164, 0;
	setp.ne.s32 	%p89, %r128, %r164;
	@!%p89 bra 	$Lt_37_707074;
	.loc	20	100	0
	ld.const.f32 	%f152, [kYCbCrOffset+0];
	bra.uni 	$Lt_37_706818;
$Lt_37_707074:
	ld.const.f32 	%f152, [kYCbCrFullRangeOffset+0];
$Lt_37_706818:
	.loc	20	107	0
	@!%p89 bra 	$Lt_37_707586;
	.loc	20	100	0
	ld.const.f32 	%f153, [kYCbCrOffset+4];
	bra.uni 	$Lt_37_707330;
$Lt_37_707586:
	ld.const.f32 	%f153, [kYCbCrFullRangeOffset+4];
$Lt_37_707330:
	.loc	20	107	0
	@!%p89 bra 	$Lt_37_708098;
	.loc	20	100	0
	ld.const.f32 	%f154, [kYCbCrOffset+8];
	bra.uni 	$Lt_37_707842;
$Lt_37_708098:
	ld.const.f32 	%f154, [kYCbCrFullRangeOffset+8];
$Lt_37_707842:
	.loc	20	424	0
	mov.f32 	%f155, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f156, %f151, %f155;
	fma.rn.ftz.f32 	%f134, %f156, %f152, %f134;
	fma.rn.ftz.f32 	%f136, %f156, %f153, %f136;
	fma.rn.ftz.f32 	%f138, %f156, %f154, %f138;
$Lt_37_706306:
	.loc	20	543	0
	mov.f32 	%f24, %f134;
	mov.f32 	%f25, %f136;
	mov.f32 	%f26, %f138;
	@!%p99 bra 	$Lt_37_703746;
	.loc	20	57	0
	@!%p43 bra 	$Lt_37_473858;
	.loc	20	59	0
	mov.f32 	%f132, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_169;
$Lt_37_473858:
	.loc	20	61	0
	mov.u32 	%r165, 64;
	setp.ne.s32 	%p112, %r21, %r165;
	@%p112 bra 	$Lt_37_474114;
	.loc	20	63	0
	mov.f32 	%f132, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_169;
$Lt_37_474114:
	.loc	20	65	0
	mov.u32 	%r166, 128;
	setp.ne.s32 	%p113, %r21, %r166;
	@%p113 bra 	$Lt_37_474370;
	.loc	20	68	0
	mov.f32 	%f132, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_169;
$Lt_37_474370:
	.loc	20	70	0
	mov.u32 	%r167, 192;
	setp.ne.s32 	%p114, %r21, %r167;
	@%p114 bra 	$Lt_37_474626;
	.loc	20	72	0
	mov.f32 	%f132, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_169;
$Lt_37_474626:
	.loc	20	76	0
	mov.f32 	%f132, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_169:
	.loc	20	546	0
	mov.f32 	%f22, %f132;
	bra.uni 	$Lt_37_703746;
$Lt_37_704002:
	.loc	20	433	0
	mov.f32 	%f157, %f22;
	mov.f32 	%f158, %f24;
	mov.f32 	%f159, %f158;
	mov.f32 	%f160, %f25;
	mov.f32 	%f161, %f160;
	mov.f32 	%f162, %f26;
	mov.f32 	%f163, %f162;
	.loc	20	435	0
	mov.u32 	%r168, 0;
	setp.eq.s32 	%p115, %r26, %r168;
	@%p115 bra 	$Lt_37_708866;
	.loc	20	57	0
	mov.u32 	%r169, 0;
	setp.ne.s32 	%p116, %r21, %r169;
	@%p116 bra 	$Lt_37_475138;
	.loc	20	59	0
	mov.f32 	%f164, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_167;
$Lt_37_475138:
	.loc	20	61	0
	mov.u32 	%r170, 64;
	setp.ne.s32 	%p117, %r21, %r170;
	@%p117 bra 	$Lt_37_475394;
	.loc	20	63	0
	mov.f32 	%f164, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_167;
$Lt_37_475394:
	.loc	20	65	0
	mov.u32 	%r171, 128;
	setp.ne.s32 	%p118, %r21, %r171;
	@%p118 bra 	$Lt_37_475650;
	.loc	20	68	0
	mov.f32 	%f164, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_167;
$Lt_37_475650:
	.loc	20	70	0
	mov.u32 	%r172, 192;
	setp.ne.s32 	%p119, %r21, %r172;
	@%p119 bra 	$Lt_37_475906;
	.loc	20	72	0
	mov.f32 	%f164, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_167;
$Lt_37_475906:
	.loc	20	76	0
	mov.f32 	%f164, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_167:
	.loc	20	118	0
	and.b32 	%r128, %r4, 2048;
	mov.s32 	%r173, 0;
	setp.ne.s32 	%p89, %r128, %r173;
	@!%p89 bra 	$Lt_37_709634;
	.loc	20	100	0
	ld.const.f32 	%f165, [kYCbCrOffset+0];
	bra.uni 	$Lt_37_709378;
$Lt_37_709634:
	ld.const.f32 	%f165, [kYCbCrFullRangeOffset+0];
$Lt_37_709378:
	.loc	20	118	0
	@!%p89 bra 	$Lt_37_710146;
	.loc	20	100	0
	ld.const.f32 	%f166, [kYCbCrOffset+4];
	bra.uni 	$Lt_37_709890;
$Lt_37_710146:
	ld.const.f32 	%f166, [kYCbCrFullRangeOffset+4];
$Lt_37_709890:
	.loc	20	118	0
	@!%p89 bra 	$Lt_37_710658;
	.loc	20	100	0
	ld.const.f32 	%f167, [kYCbCrOffset+8];
	bra.uni 	$Lt_37_710402;
$Lt_37_710658:
	ld.const.f32 	%f167, [kYCbCrFullRangeOffset+8];
$Lt_37_710402:
	.loc	20	437	0
	mov.f32 	%f168, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f169, %f164, %f168;
	mul.ftz.f32 	%f170, %f169, %f165;
	sub.ftz.f32 	%f159, %f158, %f170;
	mul.ftz.f32 	%f171, %f169, %f166;
	sub.ftz.f32 	%f161, %f160, %f171;
	mul.ftz.f32 	%f172, %f169, %f167;
	sub.ftz.f32 	%f163, %f162, %f172;
$Lt_37_708866:
	mov.f32 	%f173, 0fb70637bd;   	// -8e-006
	add.ftz.f32 	%f174, %f22, %f173;
	mov.f32 	%f175, 0f00000000;   	// 0
	setp.le.ftz.f32 	%p120, %f174, %f175;
	@!%p120 bra 	$Lt_37_711170;
	mov.f32 	%f163, 0f00000000;   	// 0
	mov.f32 	%f161, 0f00000000;   	// 0
	mov.f32 	%f159, 0f00000000;   	// 0
	mov.f32 	%f157, 0f00000000;   	// 0
	bra.uni 	$Lt_37_710914;
$Lt_37_711170:
	.loc	20	57	0
	mov.u32 	%r174, 0;
	setp.ne.s32 	%p121, %r21, %r174;
	@%p121 bra 	$Lt_37_478210;
	.loc	20	59	0
	mov.f32 	%f176, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_163;
$Lt_37_478210:
	.loc	20	61	0
	mov.u32 	%r175, 64;
	setp.ne.s32 	%p122, %r21, %r175;
	@%p122 bra 	$Lt_37_478466;
	.loc	20	63	0
	mov.f32 	%f176, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_163;
$Lt_37_478466:
	.loc	20	65	0
	mov.u32 	%r176, 128;
	setp.ne.s32 	%p123, %r21, %r176;
	@%p123 bra 	$Lt_37_478722;
	.loc	20	68	0
	mov.f32 	%f176, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_163;
$Lt_37_478722:
	.loc	20	70	0
	mov.u32 	%r177, 192;
	setp.ne.s32 	%p124, %r21, %r177;
	@%p124 bra 	$Lt_37_478978;
	.loc	20	72	0
	mov.f32 	%f176, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_163;
$Lt_37_478978:
	.loc	20	76	0
	mov.f32 	%f176, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_163:
	.loc	20	447	0
	div.approx.ftz.f32 	%f177, %f176, %f22;
	mul.ftz.f32 	%f159, %f177, %f159;
	.loc	20	448	0
	mul.ftz.f32 	%f161, %f177, %f161;
	.loc	20	449	0
	mul.ftz.f32 	%f163, %f177, %f163;
$Lt_37_710914:
	.loc	20	452	0
	mov.u32 	%r178, 0;
	setp.eq.s32 	%p125, %r26, %r178;
	@%p125 bra 	$Lt_37_711426;
	.loc	20	57	0
	mov.u32 	%r179, 0;
	setp.ne.s32 	%p126, %r21, %r179;
	@%p126 bra 	$Lt_37_479490;
	.loc	20	59	0
	mov.f32 	%f178, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_161;
$Lt_37_479490:
	.loc	20	61	0
	mov.u32 	%r180, 64;
	setp.ne.s32 	%p127, %r21, %r180;
	@%p127 bra 	$Lt_37_479746;
	.loc	20	63	0
	mov.f32 	%f178, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_161;
$Lt_37_479746:
	.loc	20	65	0
	mov.u32 	%r181, 128;
	setp.ne.s32 	%p128, %r21, %r181;
	@%p128 bra 	$Lt_37_480002;
	.loc	20	68	0
	mov.f32 	%f178, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_161;
$Lt_37_480002:
	.loc	20	70	0
	mov.u32 	%r182, 192;
	setp.ne.s32 	%p129, %r21, %r182;
	@%p129 bra 	$Lt_37_480258;
	.loc	20	72	0
	mov.f32 	%f178, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_161;
$Lt_37_480258:
	.loc	20	76	0
	mov.f32 	%f178, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_161:
	.loc	20	107	0
	and.b32 	%r128, %r4, 2048;
	mov.s32 	%r183, 0;
	setp.ne.s32 	%p89, %r128, %r183;
	@!%p89 bra 	$Lt_37_712194;
	.loc	20	100	0
	ld.const.f32 	%f179, [kYCbCrOffset+0];
	bra.uni 	$Lt_37_711938;
$Lt_37_712194:
	ld.const.f32 	%f179, [kYCbCrFullRangeOffset+0];
$Lt_37_711938:
	.loc	20	107	0
	@!%p89 bra 	$Lt_37_712706;
	.loc	20	100	0
	ld.const.f32 	%f180, [kYCbCrOffset+4];
	bra.uni 	$Lt_37_712450;
$Lt_37_712706:
	ld.const.f32 	%f180, [kYCbCrFullRangeOffset+4];
$Lt_37_712450:
	.loc	20	107	0
	@!%p89 bra 	$Lt_37_713218;
	.loc	20	100	0
	ld.const.f32 	%f181, [kYCbCrOffset+8];
	bra.uni 	$Lt_37_712962;
$Lt_37_713218:
	ld.const.f32 	%f181, [kYCbCrFullRangeOffset+8];
$Lt_37_712962:
	.loc	20	454	0
	mov.f32 	%f182, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f183, %f178, %f182;
	fma.rn.ftz.f32 	%f159, %f183, %f179, %f159;
	fma.rn.ftz.f32 	%f161, %f183, %f180, %f161;
	fma.rn.ftz.f32 	%f163, %f183, %f181, %f163;
$Lt_37_711426:
	.loc	20	551	0
	mov.f32 	%f22, %f157;
	mov.f32 	%f24, %f159;
	mov.f32 	%f25, %f161;
	mov.f32 	%f26, %f163;
$Lt_37_703746:
$L_37_678914:
$Lt_37_703234:
	.loc	20	540	0
	and.b32 	%r184, %r4, 4096;
	mov.s32 	%r185, 0;
	setp.eq.s32 	%p130, %r184, %r185;
	@!%p130 bra 	$Lt_37_713474;
	.loc	21	268	0
	mov.f32 	%f184, %f24;
	.loc	21	269	0
	mov.f32 	%f185, %f22;
	.loc	20	558	0
	mov.f32 	%f22, %f26;
	mov.f32 	%f24, %f25;
	mov.f32 	%f25, %f184;
	mov.f32 	%f26, %f185;
$Lt_37_713474:
	@!%p8 bra 	$Lt_37_713986;
	.loc	20	57	0
	mov.u32 	%r186, 0;
	setp.ne.s32 	%p131, %r21, %r186;
	@%p131 bra 	$Lt_37_482818;
	.loc	20	59	0
	mov.f32 	%f186, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_157;
$Lt_37_482818:
	.loc	20	61	0
	mov.u32 	%r187, 64;
	setp.ne.s32 	%p132, %r21, %r187;
	@%p132 bra 	$Lt_37_483074;
	.loc	20	63	0
	mov.f32 	%f186, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_157;
$Lt_37_483074:
	.loc	20	65	0
	mov.u32 	%r188, 128;
	setp.ne.s32 	%p133, %r21, %r188;
	@%p133 bra 	$Lt_37_483330;
	.loc	20	68	0
	mov.f32 	%f186, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_157;
$Lt_37_483330:
	.loc	20	70	0
	mov.u32 	%r189, 192;
	setp.ne.s32 	%p134, %r21, %r189;
	@%p134 bra 	$Lt_37_483586;
	.loc	20	72	0
	mov.f32 	%f186, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_157;
$Lt_37_483586:
	.loc	20	76	0
	mov.f32 	%f186, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_157:
	.loc	20	564	0
	mov.f32 	%f187, 0f3f000000;   	// 0.5
	add.ftz.f32 	%f188, %f22, %f187;
	mov.f32 	%f189, 0f00000000;   	// 0
	mov.f32 	%f190, 0f00000000;   	// 0
	setp.gt.ftz.f32 	%p135, %f188, %f190;
	selp.f32 	%f191, %f188, %f189, %p135;
	min.ftz.f32 	%f22, %f191, %f186;
	mov.f32 	%f192, 0f3f000000;   	// 0.5
	add.ftz.f32 	%f193, %f24, %f192;
	mov.f32 	%f194, 0f00000000;   	// 0
	mov.f32 	%f195, 0f00000000;   	// 0
	setp.gt.ftz.f32 	%p136, %f193, %f195;
	selp.f32 	%f196, %f193, %f194, %p136;
	min.ftz.f32 	%f24, %f196, %f186;
	mov.f32 	%f197, 0f3f000000;   	// 0.5
	add.ftz.f32 	%f198, %f25, %f197;
	mov.f32 	%f199, 0f00000000;   	// 0
	mov.f32 	%f200, 0f00000000;   	// 0
	setp.gt.ftz.f32 	%p137, %f198, %f200;
	selp.f32 	%f201, %f198, %f199, %p137;
	min.ftz.f32 	%f25, %f201, %f186;
	mov.f32 	%f202, 0f3f000000;   	// 0.5
	add.ftz.f32 	%f203, %f26, %f202;
	mov.f32 	%f204, 0f00000000;   	// 0
	mov.f32 	%f205, 0f00000000;   	// 0
	setp.gt.ftz.f32 	%p138, %f203, %f205;
	selp.f32 	%f206, %f203, %f204, %p138;
	min.ftz.f32 	%f26, %f206, %f186;
$Lt_37_713986:
	st.v4.f32 	[%rd2+0], {%f22,%f24,%f25,%f26};
	.loc	22	60	0
	@!%p1 bra 	$Lt_37_714754;
	.loc	22	67	0
	add.ftz.f32 	%f207, %f4, %f12;
	cvt.rzi.ftz.u32.f32 	%r190, %f207;
	shr.u32 	%r191, %r190, 1;
	cvt.rn.f32.u32 	%f208, %r191;
	.loc	20	57	0
	@!%p3 bra 	$Lt_37_485634;
	.loc	20	59	0
	mov.f32 	%f20, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_155;
$Lt_37_485634:
	.loc	20	61	0
	mov.u32 	%r192, 64;
	setp.ne.s32 	%p139, %r14, %r192;
	@%p139 bra 	$Lt_37_485890;
	.loc	20	63	0
	mov.f32 	%f20, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_155;
$Lt_37_485890:
	.loc	20	65	0
	mov.u32 	%r193, 128;
	setp.ne.s32 	%p140, %r14, %r193;
	@%p140 bra 	$Lt_37_486146;
	.loc	20	68	0
	mov.f32 	%f20, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_155;
$Lt_37_486146:
	.loc	20	70	0
	mov.u32 	%r194, 192;
	setp.ne.s32 	%p141, %r14, %r194;
	@%p141 bra 	$Lt_37_486402;
	.loc	20	72	0
	mov.f32 	%f20, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_155;
$Lt_37_486402:
	.loc	20	76	0
	mov.f32 	%f20, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_155:
	.loc	20	469	0
	add.ftz.f32 	%f209, %f8, %f16;
	cvt.rzi.ftz.u32.f32 	%r195, %f209;
	shr.u32 	%r196, %r195, 1;
	cvt.rn.f32.u32 	%f210, %r196;
	mov.f32 	%f22, %f210;
	mov.f32 	%f24, %f208;
	mov.f32 	%f25, %f6;
	mov.f32 	%f26, %f20;
	@!%p7 bra 	$Lt_37_715522;
	.loc	20	473	0
	mov.f32 	%f22, %f26;
	mov.f32 	%f24, %f6;
	mov.f32 	%f25, %f208;
	mov.f32 	%f26, %f210;
$Lt_37_715522:
	.loc	20	476	0
	mov.u32 	%r197, 0;
	setp.eq.s32 	%p142, %r31, %r197;
	@%p142 bra 	$Lt_37_775426;
	mov.u32 	%r198, 0;
	setp.ne.s32 	%p143, %r26, %r198;
	@%p143 bra 	$Lt_37_775426;
	.loc	20	57	0
	@!%p3 bra 	$Lt_37_486914;
	.loc	20	59	0
	mov.f32 	%f27, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_153;
$Lt_37_486914:
	.loc	20	61	0
	mov.u32 	%r199, 64;
	setp.ne.s32 	%p144, %r14, %r199;
	@%p144 bra 	$Lt_37_487170;
	.loc	20	63	0
	mov.f32 	%f27, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_153;
$Lt_37_487170:
	.loc	20	65	0
	mov.u32 	%r200, 128;
	setp.ne.s32 	%p145, %r14, %r200;
	@%p145 bra 	$Lt_37_487426;
	.loc	20	68	0
	mov.f32 	%f27, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_153;
$Lt_37_487426:
	.loc	20	70	0
	mov.u32 	%r201, 192;
	setp.ne.s32 	%p146, %r14, %r201;
	@%p146 bra 	$Lt_37_487682;
	.loc	20	72	0
	mov.f32 	%f27, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_153;
$Lt_37_487682:
	.loc	20	76	0
	mov.f32 	%f27, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_153:
	.loc	20	118	0
	and.b32 	%r37, %r13, 2048;
	mov.s32 	%r202, 0;
	setp.ne.s32 	%p15, %r37, %r202;
	@!%p15 bra 	$Lt_37_716290;
	.loc	20	100	0
	ld.const.f32 	%f211, [kYCbCrOffset+0];
	bra.uni 	$Lt_37_716034;
$Lt_37_716290:
	ld.const.f32 	%f211, [kYCbCrFullRangeOffset+0];
$Lt_37_716034:
	.loc	20	118	0
	@!%p15 bra 	$Lt_37_716802;
	.loc	20	100	0
	ld.const.f32 	%f212, [kYCbCrOffset+4];
	bra.uni 	$Lt_37_716546;
$Lt_37_716802:
	ld.const.f32 	%f212, [kYCbCrFullRangeOffset+4];
$Lt_37_716546:
	.loc	20	118	0
	@!%p15 bra 	$Lt_37_717314;
	.loc	20	100	0
	ld.const.f32 	%f213, [kYCbCrOffset+8];
	bra.uni 	$Lt_37_717058;
$Lt_37_717314:
	ld.const.f32 	%f213, [kYCbCrFullRangeOffset+8];
$Lt_37_717058:
	.loc	20	478	0
	mov.f32 	%f214, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f215, %f27, %f214;
	mul.ftz.f32 	%f216, %f215, %f211;
	sub.ftz.f32 	%f24, %f24, %f216;
	mul.ftz.f32 	%f217, %f215, %f212;
	sub.ftz.f32 	%f25, %f25, %f217;
	mul.ftz.f32 	%f218, %f215, %f213;
	sub.ftz.f32 	%f26, %f26, %f218;
$Lt_37_775426:
$Lt_37_193538:
	.loc	20	481	0
	@!%p16 bra 	$Lt_37_775938;
	mov.u32 	%r203, 0;
	setp.ne.s32 	%p147, %r39, %r203;
	@%p147 bra 	$Lt_37_775938;
	.loc	20	483	0
	mov.f32 	%f219, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p148, %f24, %f219;
	@!%p148 bra 	$Lt_37_717570;
	.loc	20	372	0
	neg.ftz.f32 	%f220, %f24;
	lg2.approx.ftz.f32 	%f221, %f220;
	mov.f32 	%f222, 0f3ee66666;   	// 0.45
	mul.ftz.f32 	%f223, %f221, %f222;
	ex2.approx.ftz.f32 	%f224, %f223;
	neg.ftz.f32 	%f42, %f224;
	bra.uni 	$LDWendi___log2f_215_149;
$Lt_37_717570:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f225, %f24;
	mov.f32 	%f226, 0f3ee66666;   	// 0.45
	mul.ftz.f32 	%f227, %f225, %f226;
	ex2.approx.ftz.f32 	%f42, %f227;
$LDWendi___log2f_215_149:
	.loc	20	483	0
	mov.f32 	%f228, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p149, %f25, %f228;
	@!%p149 bra 	$Lt_37_718082;
	.loc	20	372	0
	neg.ftz.f32 	%f229, %f25;
	lg2.approx.ftz.f32 	%f230, %f229;
	mov.f32 	%f231, 0f3ee66666;   	// 0.45
	mul.ftz.f32 	%f232, %f230, %f231;
	ex2.approx.ftz.f32 	%f233, %f232;
	neg.ftz.f32 	%f52, %f233;
	bra.uni 	$LDWendi___log2f_215_147;
$Lt_37_718082:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f234, %f25;
	mov.f32 	%f235, 0f3ee66666;   	// 0.45
	mul.ftz.f32 	%f236, %f234, %f235;
	ex2.approx.ftz.f32 	%f52, %f236;
$LDWendi___log2f_215_147:
	.loc	20	483	0
	mov.f32 	%f237, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p150, %f26, %f237;
	@!%p150 bra 	$Lt_37_718594;
	.loc	20	372	0
	neg.ftz.f32 	%f238, %f26;
	lg2.approx.ftz.f32 	%f239, %f238;
	mov.f32 	%f240, 0f3ee66666;   	// 0.45
	mul.ftz.f32 	%f241, %f239, %f240;
	ex2.approx.ftz.f32 	%f242, %f241;
	neg.ftz.f32 	%f62, %f242;
	bra.uni 	$LDWendi___log2f_215_145;
$Lt_37_718594:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f243, %f26;
	mov.f32 	%f244, 0f3ee66666;   	// 0.45
	mul.ftz.f32 	%f245, %f243, %f244;
	ex2.approx.ftz.f32 	%f62, %f245;
$LDWendi___log2f_215_145:
	.loc	20	483	0
	mov.f32 	%f24, %f42;
	mov.f32 	%f25, %f52;
	mov.f32 	%f26, %f62;
$Lt_37_775938:
$Lt_37_197378:
	.loc	20	486	0
	mov.u32 	%r204, 0;
	setp.ne.s32 	%p151, %r49, %r204;
	@%p151 bra 	$Lt_37_248066;
	setp.eq.s32 	%p152, %r26, %r31;
	@%p152 bra 	$Lt_37_248322;
$Lt_37_248066:
	.loc	20	490	0
	mov.u32 	%r205, 0;
	setp.ne.s32 	%p153, %r45, %r205;
	@%p153 bra 	$Lt_37_719362;
	mov.s32 	%r206, 256;
	setp.eq.s32 	%p26, %r21, %r206;
	mov.u32 	%r207, 256;
	setp.ne.s32 	%p154, %r14, %r207;
	@%p154 bra 	$Lt_37_719874;
	.loc	20	137	0
	mov.s32 	%r208, 512;
	setp.eq.s32 	%p28, %r44, %r208;
	@!%p26 bra 	$Lt_37_491010;
	.loc	20	139	0
	@!%p28 bra 	$Lt_37_491266;
	.loc	20	141	0
	mov.s64 	%rd6, %rd5;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_143;
$Lt_37_491266:
	.loc	20	145	0
	mov.s64 	%rd6, %rd7;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_143;
$Lt_37_491010:
	.loc	20	150	0
	@!%p28 bra 	$Lt_37_491522;
	.loc	20	152	0
	mov.s64 	%rd6, %rd8;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_143;
$Lt_37_491522:
	.loc	20	154	0
	and.b32 	%r209, %r4, 2048;
	mov.u32 	%r210, 0;
	setp.ne.s32 	%p155, %r209, %r210;
	@%p155 bra 	$Lt_37_491778;
	.loc	20	156	0
	mov.s64 	%rd6, %rd9;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_143;
$Lt_37_491778:
	.loc	20	160	0
	mov.s64 	%rd6, %rd10;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_143;
$Lt_37_719874:
	@!%p26 bra 	$Lt_37_720386;
	bra.uni 	$Lt_37_719106;
$Lt_37_720386:
	.loc	20	179	0
	mov.u32 	%r211, 512;
	setp.ne.s32 	%p156, %r44, %r211;
	@%p156 bra 	$Lt_37_492546;
	.loc	20	181	0
	mov.s64 	%rd6, %rd11;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_143;
$Lt_37_492546:
	.loc	20	183	0
	and.b32 	%r212, %r4, 2048;
	mov.u32 	%r213, 0;
	setp.ne.s32 	%p157, %r212, %r213;
	@%p157 bra 	$Lt_37_492802;
	.loc	20	185	0
	mov.s64 	%rd6, %rd12;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_143;
$Lt_37_492802:
	.loc	20	189	0
	mov.s64 	%rd6, %rd13;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_143;
$Lt_37_719362:
	mov.s32 	%r214, 0;
	setp.eq.s32 	%p32, %r43, %r214;
	mov.u32 	%r215, 512;
	setp.ne.s32 	%p158, %r46, %r215;
	@%p158 bra 	$Lt_37_720898;
	mov.s32 	%r216, 256;
	setp.eq.s32 	%p34, %r14, %r216;
	@!%p32 bra 	$Lt_37_721410;
	mov.s32 	%r217, 256;
	setp.eq.s32 	%p26, %r21, %r217;
	@!%p34 bra 	$Lt_37_721922;
	@!%p26 bra 	$Lt_37_719106;
	.loc	20	202	0
	mov.s64 	%rd6, %rd14;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_143;
$Lt_37_721922:
	.loc	20	211	0
	@!%p26 bra 	$Lt_37_494082;
	.loc	20	213	0
	mov.s64 	%rd6, %rd15;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_143;
$Lt_37_494082:
	.loc	20	217	0
	mov.s64 	%rd6, %rd16;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_143;
$Lt_37_721410:
	@!%p34 bra 	$Lt_37_722946;
	bra.uni 	$Lt_37_719106;
$Lt_37_722946:
	mov.s32 	%r218, 256;
	set.eq.u32.s32 	%r219, %r21, %r218;
	neg.s32 	%r220, %r219;
	and.b32 	%r221, %r4, 2048;
	mov.s32 	%r222, 0;
	set.eq.u32.s32 	%r223, %r221, %r222;
	neg.s32 	%r224, %r223;
	or.b32 	%r225, %r220, %r224;
	mov.u32 	%r226, 0;
	setp.eq.s32 	%p159, %r225, %r226;
	@%p159 bra 	$Lt_37_723458;
	bra.uni 	$Lt_37_719106;
$Lt_37_723458:
	.loc	20	250	0
	mov.s64 	%rd6, %rd17;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_143;
$Lt_37_720898:
	and.b32 	%r37, %r13, 2048;
	mov.s32 	%r227, 0;
	setp.eq.s32 	%p160, %r37, %r227;
	mov.pred 	%p37, %p160;
	mov.pred 	%p161, %p39;
	@!%p160 bra 	$Lt_37_723970;
	@!%p32 bra 	$Lt_37_719106;
	.loc	20	259	0
	mov.u32 	%r228, 256;
	setp.ne.s32 	%p162, %r21, %r228;
	@%p162 bra 	$Lt_37_495618;
	.loc	20	261	0
	mov.s64 	%rd6, %rd18;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_143;
$Lt_37_495618:
	.loc	20	265	0
	mov.s64 	%rd6, %rd19;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_143;
$Lt_37_723970:
	mov.s32 	%r229, 256;
	setp.eq.s32 	%p34, %r14, %r229;
	@!%p32 bra 	$Lt_37_724994;
	mov.s32 	%r230, 256;
	setp.eq.s32 	%p26, %r21, %r230;
	@!%p34 bra 	$Lt_37_725506;
	@!%p26 bra 	$Lt_37_719106;
	.loc	20	302	0
	mov.s64 	%rd6, %rd20;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_143;
$Lt_37_725506:
	.loc	20	311	0
	@!%p26 bra 	$Lt_37_497410;
	.loc	20	313	0
	mov.s64 	%rd6, %rd21;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_143;
$Lt_37_497410:
	.loc	20	317	0
	mov.s64 	%rd6, %rd22;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_143;
$Lt_37_724994:
	@!%p34 bra 	$Lt_37_726530;
	bra.uni 	$Lt_37_719106;
$Lt_37_726530:
	selp.s32 	%r231, 1, 0, %p37;
	mov.s32 	%r232, 256;
	set.eq.u32.s32 	%r233, %r21, %r232;
	neg.s32 	%r234, %r233;
	or.b32 	%r235, %r231, %r234;
	mov.u32 	%r236, 0;
	setp.eq.s32 	%p163, %r235, %r236;
	@%p163 bra 	$Lt_37_727042;
	bra.uni 	$Lt_37_719106;
$Lt_37_727042:
	.loc	20	350	0
	mov.s64 	%rd6, %rd23;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_143;
$Lt_37_719106:
	.loc	20	355	0
	mov.u64 	%rd6, 0;
$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_143:
	.loc	20	490	0
	ld.f32 	%f246, [%rd6+16];
	mul.ftz.f32 	%f247, %f246, %f25;
	ld.f32 	%f248, [%rd6+12];
	fma.rn.ftz.f32 	%f249, %f248, %f24, %f247;
	ld.f32 	%f250, [%rd6+20];
	fma.rn.ftz.f32 	%f251, %f250, %f26, %f249;
	ld.f32 	%f252, [%rd6+28];
	mul.ftz.f32 	%f253, %f252, %f25;
	ld.f32 	%f254, [%rd6+24];
	fma.rn.ftz.f32 	%f255, %f254, %f24, %f253;
	ld.f32 	%f256, [%rd6+32];
	fma.rn.ftz.f32 	%f257, %f256, %f26, %f255;
	ld.f32 	%f258, [%rd6+4];
	mul.ftz.f32 	%f259, %f258, %f25;
	ld.f32 	%f260, [%rd6+0];
	fma.rn.ftz.f32 	%f261, %f260, %f24, %f259;
	ld.f32 	%f262, [%rd6+8];
	fma.rn.ftz.f32 	%f24, %f262, %f26, %f261;
	mov.f32 	%f25, %f251;
	mov.f32 	%f26, %f257;
	@!%p42 bra 	$Lt_37_727810;
	.loc	20	494	0
	mov.s32 	%r237, 256;
	setp.eq.s32 	%p34, %r14, %r237;
	@!%p34 bra 	$L_37_680962;
	mov.s32 	%r238, 0;
	setp.eq.s32 	%p43, %r21, %r238;
	@%p43 bra 	$L_37_680706;
$L_37_680962:
	@!%p3 bra 	$Lt_37_777218;
	mov.u32 	%r239, 256;
	setp.ne.s32 	%p164, %r21, %r239;
	@%p164 bra 	$Lt_37_777218;
	mov.s32 	%r240, 0;
	setp.eq.s32 	%p43, %r21, %r240;
$L_37_680706:
	.loc	20	57	0
	@!%p43 bra 	$Lt_37_498946;
	.loc	20	59	0
	mov.f32 	%f83, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_141;
$Lt_37_498946:
	.loc	20	61	0
	mov.u32 	%r241, 64;
	setp.ne.s32 	%p165, %r21, %r241;
	@%p165 bra 	$Lt_37_499202;
	.loc	20	63	0
	mov.f32 	%f83, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_141;
$Lt_37_499202:
	.loc	20	65	0
	mov.u32 	%r242, 128;
	setp.ne.s32 	%p166, %r21, %r242;
	@%p166 bra 	$Lt_37_499458;
	.loc	20	68	0
	mov.f32 	%f83, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_141;
$Lt_37_499458:
	.loc	20	70	0
	mov.u32 	%r243, 192;
	setp.ne.s32 	%p167, %r21, %r243;
	@%p167 bra 	$Lt_37_499714;
	.loc	20	72	0
	mov.f32 	%f83, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_141;
$Lt_37_499714:
	.loc	20	76	0
	mov.f32 	%f83, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_141:
	.loc	20	57	0
	@!%p3 bra 	$Lt_37_499970;
	.loc	20	59	0
	mov.f32 	%f84, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_139;
$Lt_37_499970:
	.loc	20	61	0
	mov.u32 	%r244, 64;
	setp.ne.s32 	%p168, %r14, %r244;
	@%p168 bra 	$Lt_37_500226;
	.loc	20	63	0
	mov.f32 	%f84, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_139;
$Lt_37_500226:
	.loc	20	65	0
	mov.u32 	%r245, 128;
	setp.ne.s32 	%p169, %r14, %r245;
	@%p169 bra 	$Lt_37_500482;
	.loc	20	68	0
	mov.f32 	%f84, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_139;
$Lt_37_500482:
	.loc	20	70	0
	mov.u32 	%r246, 192;
	setp.ne.s32 	%p170, %r14, %r246;
	@%p170 bra 	$Lt_37_500738;
	.loc	20	72	0
	mov.f32 	%f84, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_139;
$Lt_37_500738:
	.loc	20	76	0
	mov.f32 	%f84, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_139:
	.loc	20	498	0
	div.approx.ftz.f32 	%f263, %f83, %f84;
	mul.ftz.f32 	%f22, %f22, %f263;
	bra.uni 	$Lt_37_727810;
$Lt_37_777218:
$L_37_680450:
	.loc	20	500	0
	@!%p34 bra 	$L_37_681986;
	@%p8 bra 	$L_37_681730;
$L_37_681986:
	@!%p9 bra 	$Lt_37_778242;
	mov.u32 	%r247, 256;
	setp.ne.s32 	%p171, %r21, %r247;
	@%p171 bra 	$Lt_37_778242;
$L_37_681730:
	.loc	20	57	0
	mov.u32 	%r248, 0;
	setp.ne.s32 	%p172, %r21, %r248;
	@%p172 bra 	$Lt_37_501250;
	.loc	20	59	0
	mov.f32 	%f83, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_137;
$Lt_37_501250:
	.loc	20	61	0
	mov.u32 	%r249, 64;
	setp.ne.s32 	%p173, %r21, %r249;
	@%p173 bra 	$Lt_37_501506;
	.loc	20	63	0
	mov.f32 	%f83, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_137;
$Lt_37_501506:
	.loc	20	65	0
	mov.u32 	%r250, 128;
	setp.ne.s32 	%p174, %r21, %r250;
	@%p174 bra 	$Lt_37_501762;
	.loc	20	68	0
	mov.f32 	%f83, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_137;
$Lt_37_501762:
	.loc	20	70	0
	mov.u32 	%r251, 192;
	setp.ne.s32 	%p175, %r21, %r251;
	@%p175 bra 	$Lt_37_502018;
	.loc	20	72	0
	mov.f32 	%f83, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_137;
$Lt_37_502018:
	.loc	20	76	0
	mov.f32 	%f83, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_137:
	.loc	20	57	0
	@!%p3 bra 	$Lt_37_502274;
	.loc	20	59	0
	mov.f32 	%f84, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_135;
$Lt_37_502274:
	.loc	20	61	0
	mov.u32 	%r252, 64;
	setp.ne.s32 	%p176, %r14, %r252;
	@%p176 bra 	$Lt_37_502530;
	.loc	20	63	0
	mov.f32 	%f84, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_135;
$Lt_37_502530:
	.loc	20	65	0
	mov.u32 	%r253, 128;
	setp.ne.s32 	%p177, %r14, %r253;
	@%p177 bra 	$Lt_37_502786;
	.loc	20	68	0
	mov.f32 	%f84, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_135;
$Lt_37_502786:
	.loc	20	70	0
	mov.u32 	%r254, 192;
	setp.ne.s32 	%p178, %r14, %r254;
	@%p178 bra 	$Lt_37_503042;
	.loc	20	72	0
	mov.f32 	%f84, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_135;
$Lt_37_503042:
	.loc	20	76	0
	mov.f32 	%f84, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_135:
	.loc	20	504	0
	div.approx.ftz.f32 	%f264, %f83, %f84;
	mul.ftz.f32 	%f22, %f22, %f264;
	.loc	20	57	0
	@!%p3 bra 	$Lt_37_504322;
	.loc	20	59	0
	mov.f32 	%f84, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_131;
$Lt_37_504322:
	.loc	20	61	0
	mov.u32 	%r255, 64;
	setp.ne.s32 	%p179, %r14, %r255;
	@%p179 bra 	$Lt_37_504578;
	.loc	20	63	0
	mov.f32 	%f84, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_131;
$Lt_37_504578:
	.loc	20	65	0
	mov.u32 	%r256, 128;
	setp.ne.s32 	%p180, %r14, %r256;
	@%p180 bra 	$Lt_37_504834;
	.loc	20	68	0
	mov.f32 	%f84, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_131;
$Lt_37_504834:
	.loc	20	70	0
	mov.u32 	%r257, 192;
	setp.ne.s32 	%p181, %r14, %r257;
	@%p181 bra 	$Lt_37_505090;
	.loc	20	72	0
	mov.f32 	%f84, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_131;
$Lt_37_505090:
	.loc	20	76	0
	mov.f32 	%f84, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_131:
	.loc	20	505	0
	mov.f32 	%f265, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f266, %f265, %f84;
	mul.ftz.f32 	%f24, %f24, %f266;
	.loc	20	57	0
	@!%p3 bra 	$Lt_37_506370;
	.loc	20	59	0
	mov.f32 	%f84, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_127;
$Lt_37_506370:
	.loc	20	61	0
	mov.u32 	%r258, 64;
	setp.ne.s32 	%p182, %r14, %r258;
	@%p182 bra 	$Lt_37_506626;
	.loc	20	63	0
	mov.f32 	%f84, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_127;
$Lt_37_506626:
	.loc	20	65	0
	mov.u32 	%r259, 128;
	setp.ne.s32 	%p183, %r14, %r259;
	@%p183 bra 	$Lt_37_506882;
	.loc	20	68	0
	mov.f32 	%f84, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_127;
$Lt_37_506882:
	.loc	20	70	0
	mov.u32 	%r260, 192;
	setp.ne.s32 	%p184, %r14, %r260;
	@%p184 bra 	$Lt_37_507138;
	.loc	20	72	0
	mov.f32 	%f84, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_127;
$Lt_37_507138:
	.loc	20	76	0
	mov.f32 	%f84, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_127:
	.loc	20	506	0
	mov.f32 	%f267, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f268, %f267, %f84;
	mul.ftz.f32 	%f25, %f251, %f268;
	.loc	20	57	0
	@!%p3 bra 	$Lt_37_508418;
	.loc	20	59	0
	mov.f32 	%f84, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_123;
$Lt_37_508418:
	.loc	20	61	0
	mov.u32 	%r261, 64;
	setp.ne.s32 	%p185, %r14, %r261;
	@%p185 bra 	$Lt_37_508674;
	.loc	20	63	0
	mov.f32 	%f84, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_123;
$Lt_37_508674:
	.loc	20	65	0
	mov.u32 	%r262, 128;
	setp.ne.s32 	%p186, %r14, %r262;
	@%p186 bra 	$Lt_37_508930;
	.loc	20	68	0
	mov.f32 	%f84, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_123;
$Lt_37_508930:
	.loc	20	70	0
	mov.u32 	%r263, 192;
	setp.ne.s32 	%p187, %r14, %r263;
	@%p187 bra 	$Lt_37_509186;
	.loc	20	72	0
	mov.f32 	%f84, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_123;
$Lt_37_509186:
	.loc	20	76	0
	mov.f32 	%f84, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_123:
	.loc	20	507	0
	mov.f32 	%f269, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f270, %f269, %f84;
	mul.ftz.f32 	%f26, %f257, %f270;
	bra.uni 	$Lt_37_727810;
$Lt_37_778242:
$L_37_681474:
	.loc	20	57	0
	@!%p3 bra 	$Lt_37_509442;
	.loc	20	59	0
	mov.f32 	%f83, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_121;
$Lt_37_509442:
	.loc	20	61	0
	mov.u32 	%r264, 64;
	setp.ne.s32 	%p188, %r14, %r264;
	@%p188 bra 	$Lt_37_509698;
	.loc	20	63	0
	mov.f32 	%f83, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_121;
$Lt_37_509698:
	.loc	20	65	0
	mov.u32 	%r265, 128;
	setp.ne.s32 	%p189, %r14, %r265;
	@%p189 bra 	$Lt_37_509954;
	.loc	20	68	0
	mov.f32 	%f83, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_121;
$Lt_37_509954:
	.loc	20	70	0
	mov.u32 	%r266, 192;
	setp.ne.s32 	%p190, %r14, %r266;
	@%p190 bra 	$Lt_37_510210;
	.loc	20	72	0
	mov.f32 	%f83, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_121;
$Lt_37_510210:
	.loc	20	76	0
	mov.f32 	%f83, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_121:
	.loc	20	511	0
	mov.f32 	%f271, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f272, %f83, %f271;
	mul.ftz.f32 	%f22, %f272, %f22;
	mul.ftz.f32 	%f24, %f272, %f24;
	mul.ftz.f32 	%f25, %f272, %f251;
	mul.ftz.f32 	%f26, %f272, %f257;
	bra.uni 	$Lt_37_727810;
$Lt_37_248322:
	.loc	20	486	0
	@!%p42 bra 	$Lt_37_727810;
	.loc	20	57	0
	mov.u32 	%r267, 0;
	setp.ne.s32 	%p191, %r21, %r267;
	@%p191 bra 	$Lt_37_511746;
	.loc	20	59	0
	mov.f32 	%f83, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_117;
$Lt_37_511746:
	.loc	20	61	0
	mov.u32 	%r268, 64;
	setp.ne.s32 	%p192, %r21, %r268;
	@%p192 bra 	$Lt_37_512002;
	.loc	20	63	0
	mov.f32 	%f83, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_117;
$Lt_37_512002:
	.loc	20	65	0
	mov.u32 	%r269, 128;
	setp.ne.s32 	%p193, %r21, %r269;
	@%p193 bra 	$Lt_37_512258;
	.loc	20	68	0
	mov.f32 	%f83, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_117;
$Lt_37_512258:
	.loc	20	70	0
	mov.u32 	%r270, 192;
	setp.ne.s32 	%p194, %r21, %r270;
	@%p194 bra 	$Lt_37_512514;
	.loc	20	72	0
	mov.f32 	%f83, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_117;
$Lt_37_512514:
	.loc	20	76	0
	mov.f32 	%f83, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_117:
	.loc	20	57	0
	@!%p3 bra 	$Lt_37_512770;
	.loc	20	59	0
	mov.f32 	%f84, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_115;
$Lt_37_512770:
	.loc	20	61	0
	mov.u32 	%r271, 64;
	setp.ne.s32 	%p195, %r14, %r271;
	@%p195 bra 	$Lt_37_513026;
	.loc	20	63	0
	mov.f32 	%f84, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_115;
$Lt_37_513026:
	.loc	20	65	0
	mov.u32 	%r272, 128;
	setp.ne.s32 	%p196, %r14, %r272;
	@%p196 bra 	$Lt_37_513282;
	.loc	20	68	0
	mov.f32 	%f84, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_115;
$Lt_37_513282:
	.loc	20	70	0
	mov.u32 	%r273, 192;
	setp.ne.s32 	%p197, %r14, %r273;
	@%p197 bra 	$Lt_37_513538;
	.loc	20	72	0
	mov.f32 	%f84, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_115;
$Lt_37_513538:
	.loc	20	76	0
	mov.f32 	%f84, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_115:
	.loc	20	517	0
	div.approx.ftz.f32 	%f273, %f83, %f84;
	mul.ftz.f32 	%f22, %f273, %f22;
	mul.ftz.f32 	%f24, %f273, %f24;
	mul.ftz.f32 	%f25, %f273, %f25;
	mul.ftz.f32 	%f26, %f273, %f26;
$Lt_37_727810:
$Lt_37_250626:
	.loc	20	520	0
	@!%p78 bra 	$Lt_37_778754;
	mov.u32 	%r274, 0;
	setp.ne.s32 	%p198, %r40, %r274;
	@%p198 bra 	$Lt_37_778754;
	.loc	20	522	0
	mov.f32 	%f274, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p199, %f24, %f274;
	@!%p199 bra 	$Lt_37_728322;
	.loc	20	372	0
	neg.ftz.f32 	%f275, %f24;
	lg2.approx.ftz.f32 	%f276, %f275;
	mov.f32 	%f277, 0f400e38e4;   	// 2.22222
	mul.ftz.f32 	%f278, %f276, %f277;
	ex2.approx.ftz.f32 	%f279, %f278;
	neg.ftz.f32 	%f102, %f279;
	bra.uni 	$LDWendi___log2f_215_113;
$Lt_37_728322:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f280, %f24;
	mov.f32 	%f281, 0f400e38e4;   	// 2.22222
	mul.ftz.f32 	%f282, %f280, %f281;
	ex2.approx.ftz.f32 	%f102, %f282;
$LDWendi___log2f_215_113:
	.loc	20	522	0
	mov.f32 	%f283, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p200, %f25, %f283;
	@!%p200 bra 	$Lt_37_728834;
	.loc	20	372	0
	neg.ftz.f32 	%f284, %f25;
	lg2.approx.ftz.f32 	%f285, %f284;
	mov.f32 	%f286, 0f400e38e4;   	// 2.22222
	mul.ftz.f32 	%f287, %f285, %f286;
	ex2.approx.ftz.f32 	%f288, %f287;
	neg.ftz.f32 	%f112, %f288;
	bra.uni 	$LDWendi___log2f_215_111;
$Lt_37_728834:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f289, %f25;
	mov.f32 	%f290, 0f400e38e4;   	// 2.22222
	mul.ftz.f32 	%f291, %f289, %f290;
	ex2.approx.ftz.f32 	%f112, %f291;
$LDWendi___log2f_215_111:
	.loc	20	522	0
	mov.f32 	%f292, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p201, %f26, %f292;
	@!%p201 bra 	$Lt_37_729346;
	.loc	20	372	0
	neg.ftz.f32 	%f293, %f26;
	lg2.approx.ftz.f32 	%f294, %f293;
	mov.f32 	%f295, 0f400e38e4;   	// 2.22222
	mul.ftz.f32 	%f296, %f294, %f295;
	ex2.approx.ftz.f32 	%f297, %f296;
	neg.ftz.f32 	%f122, %f297;
	bra.uni 	$LDWendi___log2f_215_109;
$Lt_37_729346:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f298, %f26;
	mov.f32 	%f299, 0f400e38e4;   	// 2.22222
	mul.ftz.f32 	%f300, %f298, %f299;
	ex2.approx.ftz.f32 	%f122, %f300;
$LDWendi___log2f_215_109:
	.loc	20	522	0
	mov.f32 	%f24, %f102;
	mov.f32 	%f25, %f112;
	mov.f32 	%f26, %f122;
$Lt_37_778754:
$Lt_37_252674:
	.loc	20	525	0
	mov.u32 	%r275, 0;
	setp.eq.s32 	%p202, %r26, %r275;
	@%p202 bra 	$Lt_37_779266;
	mov.u32 	%r276, 0;
	setp.ne.s32 	%p203, %r31, %r276;
	@%p203 bra 	$Lt_37_779266;
	.loc	20	57	0
	mov.u32 	%r277, 0;
	setp.ne.s32 	%p204, %r21, %r277;
	@%p204 bra 	$Lt_37_514562;
	.loc	20	59	0
	mov.f32 	%f126, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_107;
$Lt_37_514562:
	.loc	20	61	0
	mov.u32 	%r278, 64;
	setp.ne.s32 	%p205, %r21, %r278;
	@%p205 bra 	$Lt_37_514818;
	.loc	20	63	0
	mov.f32 	%f126, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_107;
$Lt_37_514818:
	.loc	20	65	0
	mov.u32 	%r279, 128;
	setp.ne.s32 	%p206, %r21, %r279;
	@%p206 bra 	$Lt_37_515074;
	.loc	20	68	0
	mov.f32 	%f126, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_107;
$Lt_37_515074:
	.loc	20	70	0
	mov.u32 	%r280, 192;
	setp.ne.s32 	%p207, %r21, %r280;
	@%p207 bra 	$Lt_37_515330;
	.loc	20	72	0
	mov.f32 	%f126, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_107;
$Lt_37_515330:
	.loc	20	76	0
	mov.f32 	%f126, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_107:
	.loc	20	107	0
	and.b32 	%r128, %r4, 2048;
	mov.s32 	%r281, 0;
	setp.ne.s32 	%p89, %r128, %r281;
	@!%p89 bra 	$Lt_37_730114;
	.loc	20	100	0
	ld.const.f32 	%f301, [kYCbCrOffset+0];
	bra.uni 	$Lt_37_729858;
$Lt_37_730114:
	ld.const.f32 	%f301, [kYCbCrFullRangeOffset+0];
$Lt_37_729858:
	.loc	20	107	0
	@!%p89 bra 	$Lt_37_730626;
	.loc	20	100	0
	ld.const.f32 	%f302, [kYCbCrOffset+4];
	bra.uni 	$Lt_37_730370;
$Lt_37_730626:
	ld.const.f32 	%f302, [kYCbCrFullRangeOffset+4];
$Lt_37_730370:
	.loc	20	107	0
	@!%p89 bra 	$Lt_37_731138;
	.loc	20	100	0
	ld.const.f32 	%f303, [kYCbCrOffset+8];
	bra.uni 	$Lt_37_730882;
$Lt_37_731138:
	ld.const.f32 	%f303, [kYCbCrFullRangeOffset+8];
$Lt_37_730882:
	.loc	20	527	0
	mov.f32 	%f304, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f305, %f126, %f304;
	fma.rn.ftz.f32 	%f24, %f305, %f301, %f24;
	fma.rn.ftz.f32 	%f25, %f305, %f302, %f25;
	fma.rn.ftz.f32 	%f26, %f305, %f303, %f26;
$Lt_37_779266:
$Lt_37_259074:
	.loc	20	525	0
	@!%p90 bra 	$Lt_37_731906;
	.loc	20	532	0
	mov.u32 	%r282, 8;
	setp.ne.s32 	%p208, %r131, %r282;
	@%p208 bra 	$L_37_683266;
	mov.u32 	%r283, 12;
	setp.eq.s32 	%p209, %r130, %r283;
	@%p209 bra 	$Lt_37_780034;
$L_37_683266:
	mov.u32 	%r284, 12;
	setp.eq.s32 	%p210, %r131, %r284;
	@%p210 bra 	$Lt_37_780034;
	mov.u32 	%r285, 0;
	setp.ne.s32 	%p211, %r131, %r285;
	@%p211 bra 	$L_37_682498;
$Lt_37_780034:
$L_37_682754:
	.loc	20	57	0
	mov.u32 	%r286, 0;
	setp.ne.s32 	%p212, %r21, %r286;
	@%p212 bra 	$Lt_37_517890;
	.loc	20	59	0
	mov.f32 	%f132, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_103;
$Lt_37_517890:
	.loc	20	61	0
	mov.u32 	%r287, 64;
	setp.ne.s32 	%p213, %r21, %r287;
	@%p213 bra 	$Lt_37_518146;
	.loc	20	63	0
	mov.f32 	%f132, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_103;
$Lt_37_518146:
	.loc	20	65	0
	mov.u32 	%r288, 128;
	setp.ne.s32 	%p214, %r21, %r288;
	@%p214 bra 	$Lt_37_518402;
	.loc	20	68	0
	mov.f32 	%f132, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_103;
$Lt_37_518402:
	.loc	20	70	0
	mov.u32 	%r289, 192;
	setp.ne.s32 	%p215, %r21, %r289;
	@%p215 bra 	$Lt_37_518658;
	.loc	20	72	0
	mov.f32 	%f132, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_103;
$Lt_37_518658:
	.loc	20	76	0
	mov.f32 	%f132, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_103:
	.loc	20	536	0
	mov.f32 	%f22, %f132;
	bra.uni 	$Lt_37_731906;
$L_37_682498:
	.loc	20	540	0
	mov.s32 	%r290, 12;
	setp.eq.s32 	%p99, %r130, %r290;
	mov.s32 	%r291, 4;
	set.eq.u32.s32 	%r292, %r131, %r291;
	neg.s32 	%r293, %r292;
	selp.s32 	%r294, 1, 0, %p99;
	mov.s32 	%r295, 8;
	set.eq.u32.s32 	%r296, %r130, %r295;
	neg.s32 	%r297, %r296;
	or.b32 	%r298, %r294, %r297;
	and.b32 	%r299, %r293, %r298;
	mov.u32 	%r300, 0;
	setp.eq.s32 	%p216, %r299, %r300;
	@%p216 bra 	$Lt_37_732162;
	.loc	20	410	0
	mov.f32 	%f306, %f24;
	mov.f32 	%f134, %f306;
	mov.f32 	%f307, %f25;
	mov.f32 	%f136, %f307;
	mov.f32 	%f308, %f26;
	mov.f32 	%f138, %f308;
	.loc	20	57	0
	mov.s32 	%r301, 0;
	setp.eq.s32 	%p43, %r21, %r301;
	@!%p43 bra 	$Lt_37_519170;
	.loc	20	59	0
	mov.f32 	%f139, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_101;
$Lt_37_519170:
	.loc	20	61	0
	mov.u32 	%r302, 64;
	setp.ne.s32 	%p217, %r21, %r302;
	@%p217 bra 	$Lt_37_519426;
	.loc	20	63	0
	mov.f32 	%f139, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_101;
$Lt_37_519426:
	.loc	20	65	0
	mov.u32 	%r303, 128;
	setp.ne.s32 	%p218, %r21, %r303;
	@%p218 bra 	$Lt_37_519682;
	.loc	20	68	0
	mov.f32 	%f139, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_101;
$Lt_37_519682:
	.loc	20	70	0
	mov.u32 	%r304, 192;
	setp.ne.s32 	%p219, %r21, %r304;
	@%p219 bra 	$Lt_37_519938;
	.loc	20	72	0
	mov.f32 	%f139, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_101;
$Lt_37_519938:
	.loc	20	76	0
	mov.f32 	%f139, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_101:
	.loc	20	413	0
	mov.u32 	%r305, 0;
	setp.eq.s32 	%p220, %r26, %r305;
	@%p220 bra 	$Lt_37_732418;
	.loc	20	57	0
	@!%p43 bra 	$Lt_37_520450;
	.loc	20	59	0
	mov.f32 	%f140, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_99;
$Lt_37_520450:
	.loc	20	61	0
	mov.u32 	%r306, 64;
	setp.ne.s32 	%p221, %r21, %r306;
	@%p221 bra 	$Lt_37_520706;
	.loc	20	63	0
	mov.f32 	%f140, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_99;
$Lt_37_520706:
	.loc	20	65	0
	mov.u32 	%r307, 128;
	setp.ne.s32 	%p222, %r21, %r307;
	@%p222 bra 	$Lt_37_520962;
	.loc	20	68	0
	mov.f32 	%f140, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_99;
$Lt_37_520962:
	.loc	20	70	0
	mov.u32 	%r308, 192;
	setp.ne.s32 	%p223, %r21, %r308;
	@%p223 bra 	$Lt_37_521218;
	.loc	20	72	0
	mov.f32 	%f140, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_99;
$Lt_37_521218:
	.loc	20	76	0
	mov.f32 	%f140, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_99:
	.loc	20	118	0
	and.b32 	%r128, %r4, 2048;
	mov.s32 	%r309, 0;
	setp.ne.s32 	%p89, %r128, %r309;
	@!%p89 bra 	$Lt_37_733186;
	.loc	20	100	0
	ld.const.f32 	%f309, [kYCbCrOffset+0];
	bra.uni 	$Lt_37_732930;
$Lt_37_733186:
	ld.const.f32 	%f309, [kYCbCrFullRangeOffset+0];
$Lt_37_732930:
	.loc	20	118	0
	@!%p89 bra 	$Lt_37_733698;
	.loc	20	100	0
	ld.const.f32 	%f310, [kYCbCrOffset+4];
	bra.uni 	$Lt_37_733442;
$Lt_37_733698:
	ld.const.f32 	%f310, [kYCbCrFullRangeOffset+4];
$Lt_37_733442:
	.loc	20	118	0
	@!%p89 bra 	$Lt_37_734210;
	.loc	20	100	0
	ld.const.f32 	%f311, [kYCbCrOffset+8];
	bra.uni 	$Lt_37_733954;
$Lt_37_734210:
	ld.const.f32 	%f311, [kYCbCrFullRangeOffset+8];
$Lt_37_733954:
	.loc	20	415	0
	mov.f32 	%f312, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f313, %f140, %f312;
	mul.ftz.f32 	%f314, %f313, %f309;
	sub.ftz.f32 	%f134, %f306, %f314;
	mul.ftz.f32 	%f315, %f313, %f310;
	sub.ftz.f32 	%f136, %f307, %f315;
	mul.ftz.f32 	%f316, %f313, %f311;
	sub.ftz.f32 	%f138, %f308, %f316;
$Lt_37_732418:
	.loc	20	418	0
	rcp.approx.ftz.f32 	%f317, %f139;
	mul.ftz.f32 	%f318, %f317, %f22;
	mul.ftz.f32 	%f134, %f318, %f134;
	.loc	20	419	0
	mul.ftz.f32 	%f136, %f318, %f136;
	.loc	20	420	0
	mul.ftz.f32 	%f138, %f318, %f138;
	.loc	20	422	0
	mov.u32 	%r310, 0;
	setp.eq.s32 	%p224, %r26, %r310;
	@%p224 bra 	$Lt_37_734466;
	.loc	20	57	0
	@!%p43 bra 	$Lt_37_523522;
	.loc	20	59	0
	mov.f32 	%f151, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_95;
$Lt_37_523522:
	.loc	20	61	0
	mov.u32 	%r311, 64;
	setp.ne.s32 	%p225, %r21, %r311;
	@%p225 bra 	$Lt_37_523778;
	.loc	20	63	0
	mov.f32 	%f151, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_95;
$Lt_37_523778:
	.loc	20	65	0
	mov.u32 	%r312, 128;
	setp.ne.s32 	%p226, %r21, %r312;
	@%p226 bra 	$Lt_37_524034;
	.loc	20	68	0
	mov.f32 	%f151, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_95;
$Lt_37_524034:
	.loc	20	70	0
	mov.u32 	%r313, 192;
	setp.ne.s32 	%p227, %r21, %r313;
	@%p227 bra 	$Lt_37_524290;
	.loc	20	72	0
	mov.f32 	%f151, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_95;
$Lt_37_524290:
	.loc	20	76	0
	mov.f32 	%f151, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_95:
	.loc	20	107	0
	and.b32 	%r128, %r4, 2048;
	mov.s32 	%r314, 0;
	setp.ne.s32 	%p89, %r128, %r314;
	@!%p89 bra 	$Lt_37_735234;
	.loc	20	100	0
	ld.const.f32 	%f319, [kYCbCrOffset+0];
	bra.uni 	$Lt_37_734978;
$Lt_37_735234:
	ld.const.f32 	%f319, [kYCbCrFullRangeOffset+0];
$Lt_37_734978:
	.loc	20	107	0
	@!%p89 bra 	$Lt_37_735746;
	.loc	20	100	0
	ld.const.f32 	%f320, [kYCbCrOffset+4];
	bra.uni 	$Lt_37_735490;
$Lt_37_735746:
	ld.const.f32 	%f320, [kYCbCrFullRangeOffset+4];
$Lt_37_735490:
	.loc	20	107	0
	@!%p89 bra 	$Lt_37_736258;
	.loc	20	100	0
	ld.const.f32 	%f321, [kYCbCrOffset+8];
	bra.uni 	$Lt_37_736002;
$Lt_37_736258:
	ld.const.f32 	%f321, [kYCbCrFullRangeOffset+8];
$Lt_37_736002:
	.loc	20	424	0
	mov.f32 	%f322, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f323, %f151, %f322;
	fma.rn.ftz.f32 	%f134, %f323, %f319, %f134;
	fma.rn.ftz.f32 	%f136, %f323, %f320, %f136;
	fma.rn.ftz.f32 	%f138, %f323, %f321, %f138;
$Lt_37_734466:
	.loc	20	543	0
	mov.f32 	%f24, %f134;
	mov.f32 	%f25, %f136;
	mov.f32 	%f26, %f138;
	@!%p99 bra 	$Lt_37_731906;
	.loc	20	57	0
	@!%p43 bra 	$Lt_37_526594;
	.loc	20	59	0
	mov.f32 	%f132, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_91;
$Lt_37_526594:
	.loc	20	61	0
	mov.u32 	%r315, 64;
	setp.ne.s32 	%p228, %r21, %r315;
	@%p228 bra 	$Lt_37_526850;
	.loc	20	63	0
	mov.f32 	%f132, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_91;
$Lt_37_526850:
	.loc	20	65	0
	mov.u32 	%r316, 128;
	setp.ne.s32 	%p229, %r21, %r316;
	@%p229 bra 	$Lt_37_527106;
	.loc	20	68	0
	mov.f32 	%f132, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_91;
$Lt_37_527106:
	.loc	20	70	0
	mov.u32 	%r317, 192;
	setp.ne.s32 	%p230, %r21, %r317;
	@%p230 bra 	$Lt_37_527362;
	.loc	20	72	0
	mov.f32 	%f132, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_91;
$Lt_37_527362:
	.loc	20	76	0
	mov.f32 	%f132, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_91:
	.loc	20	546	0
	mov.f32 	%f22, %f132;
	bra.uni 	$Lt_37_731906;
$Lt_37_732162:
	.loc	20	433	0
	mov.f32 	%f157, %f22;
	mov.f32 	%f324, %f24;
	mov.f32 	%f159, %f324;
	mov.f32 	%f325, %f25;
	mov.f32 	%f161, %f325;
	mov.f32 	%f326, %f26;
	mov.f32 	%f163, %f326;
	.loc	20	435	0
	mov.u32 	%r318, 0;
	setp.eq.s32 	%p231, %r26, %r318;
	@%p231 bra 	$Lt_37_737026;
	.loc	20	57	0
	mov.u32 	%r319, 0;
	setp.ne.s32 	%p232, %r21, %r319;
	@%p232 bra 	$Lt_37_527874;
	.loc	20	59	0
	mov.f32 	%f164, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_89;
$Lt_37_527874:
	.loc	20	61	0
	mov.u32 	%r320, 64;
	setp.ne.s32 	%p233, %r21, %r320;
	@%p233 bra 	$Lt_37_528130;
	.loc	20	63	0
	mov.f32 	%f164, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_89;
$Lt_37_528130:
	.loc	20	65	0
	mov.u32 	%r321, 128;
	setp.ne.s32 	%p234, %r21, %r321;
	@%p234 bra 	$Lt_37_528386;
	.loc	20	68	0
	mov.f32 	%f164, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_89;
$Lt_37_528386:
	.loc	20	70	0
	mov.u32 	%r322, 192;
	setp.ne.s32 	%p235, %r21, %r322;
	@%p235 bra 	$Lt_37_528642;
	.loc	20	72	0
	mov.f32 	%f164, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_89;
$Lt_37_528642:
	.loc	20	76	0
	mov.f32 	%f164, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_89:
	.loc	20	118	0
	and.b32 	%r128, %r4, 2048;
	mov.s32 	%r323, 0;
	setp.ne.s32 	%p89, %r128, %r323;
	@!%p89 bra 	$Lt_37_737794;
	.loc	20	100	0
	ld.const.f32 	%f327, [kYCbCrOffset+0];
	bra.uni 	$Lt_37_737538;
$Lt_37_737794:
	ld.const.f32 	%f327, [kYCbCrFullRangeOffset+0];
$Lt_37_737538:
	.loc	20	118	0
	@!%p89 bra 	$Lt_37_738306;
	.loc	20	100	0
	ld.const.f32 	%f328, [kYCbCrOffset+4];
	bra.uni 	$Lt_37_738050;
$Lt_37_738306:
	ld.const.f32 	%f328, [kYCbCrFullRangeOffset+4];
$Lt_37_738050:
	.loc	20	118	0
	@!%p89 bra 	$Lt_37_738818;
	.loc	20	100	0
	ld.const.f32 	%f329, [kYCbCrOffset+8];
	bra.uni 	$Lt_37_738562;
$Lt_37_738818:
	ld.const.f32 	%f329, [kYCbCrFullRangeOffset+8];
$Lt_37_738562:
	.loc	20	437	0
	mov.f32 	%f330, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f331, %f164, %f330;
	mul.ftz.f32 	%f332, %f331, %f327;
	sub.ftz.f32 	%f159, %f324, %f332;
	mul.ftz.f32 	%f333, %f331, %f328;
	sub.ftz.f32 	%f161, %f325, %f333;
	mul.ftz.f32 	%f334, %f331, %f329;
	sub.ftz.f32 	%f163, %f326, %f334;
$Lt_37_737026:
	mov.f32 	%f335, 0fb70637bd;   	// -8e-006
	add.ftz.f32 	%f336, %f22, %f335;
	mov.f32 	%f337, 0f00000000;   	// 0
	setp.le.ftz.f32 	%p236, %f336, %f337;
	@!%p236 bra 	$Lt_37_739330;
	mov.f32 	%f163, 0f00000000;   	// 0
	mov.f32 	%f161, 0f00000000;   	// 0
	mov.f32 	%f159, 0f00000000;   	// 0
	mov.f32 	%f157, 0f00000000;   	// 0
	bra.uni 	$Lt_37_739074;
$Lt_37_739330:
	.loc	20	57	0
	mov.u32 	%r324, 0;
	setp.ne.s32 	%p237, %r21, %r324;
	@%p237 bra 	$Lt_37_530946;
	.loc	20	59	0
	mov.f32 	%f176, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_85;
$Lt_37_530946:
	.loc	20	61	0
	mov.u32 	%r325, 64;
	setp.ne.s32 	%p238, %r21, %r325;
	@%p238 bra 	$Lt_37_531202;
	.loc	20	63	0
	mov.f32 	%f176, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_85;
$Lt_37_531202:
	.loc	20	65	0
	mov.u32 	%r326, 128;
	setp.ne.s32 	%p239, %r21, %r326;
	@%p239 bra 	$Lt_37_531458;
	.loc	20	68	0
	mov.f32 	%f176, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_85;
$Lt_37_531458:
	.loc	20	70	0
	mov.u32 	%r327, 192;
	setp.ne.s32 	%p240, %r21, %r327;
	@%p240 bra 	$Lt_37_531714;
	.loc	20	72	0
	mov.f32 	%f176, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_85;
$Lt_37_531714:
	.loc	20	76	0
	mov.f32 	%f176, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_85:
	.loc	20	447	0
	div.approx.ftz.f32 	%f338, %f176, %f22;
	mul.ftz.f32 	%f159, %f338, %f159;
	.loc	20	448	0
	mul.ftz.f32 	%f161, %f338, %f161;
	.loc	20	449	0
	mul.ftz.f32 	%f163, %f338, %f163;
$Lt_37_739074:
	.loc	20	452	0
	mov.u32 	%r328, 0;
	setp.eq.s32 	%p241, %r26, %r328;
	@%p241 bra 	$Lt_37_739586;
	.loc	20	57	0
	mov.u32 	%r329, 0;
	setp.ne.s32 	%p242, %r21, %r329;
	@%p242 bra 	$Lt_37_532226;
	.loc	20	59	0
	mov.f32 	%f178, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_83;
$Lt_37_532226:
	.loc	20	61	0
	mov.u32 	%r330, 64;
	setp.ne.s32 	%p243, %r21, %r330;
	@%p243 bra 	$Lt_37_532482;
	.loc	20	63	0
	mov.f32 	%f178, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_83;
$Lt_37_532482:
	.loc	20	65	0
	mov.u32 	%r331, 128;
	setp.ne.s32 	%p244, %r21, %r331;
	@%p244 bra 	$Lt_37_532738;
	.loc	20	68	0
	mov.f32 	%f178, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_83;
$Lt_37_532738:
	.loc	20	70	0
	mov.u32 	%r332, 192;
	setp.ne.s32 	%p245, %r21, %r332;
	@%p245 bra 	$Lt_37_532994;
	.loc	20	72	0
	mov.f32 	%f178, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_83;
$Lt_37_532994:
	.loc	20	76	0
	mov.f32 	%f178, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_83:
	.loc	20	107	0
	and.b32 	%r128, %r4, 2048;
	mov.s32 	%r333, 0;
	setp.ne.s32 	%p89, %r128, %r333;
	@!%p89 bra 	$Lt_37_740354;
	.loc	20	100	0
	ld.const.f32 	%f339, [kYCbCrOffset+0];
	bra.uni 	$Lt_37_740098;
$Lt_37_740354:
	ld.const.f32 	%f339, [kYCbCrFullRangeOffset+0];
$Lt_37_740098:
	.loc	20	107	0
	@!%p89 bra 	$Lt_37_740866;
	.loc	20	100	0
	ld.const.f32 	%f340, [kYCbCrOffset+4];
	bra.uni 	$Lt_37_740610;
$Lt_37_740866:
	ld.const.f32 	%f340, [kYCbCrFullRangeOffset+4];
$Lt_37_740610:
	.loc	20	107	0
	@!%p89 bra 	$Lt_37_741378;
	.loc	20	100	0
	ld.const.f32 	%f341, [kYCbCrOffset+8];
	bra.uni 	$Lt_37_741122;
$Lt_37_741378:
	ld.const.f32 	%f341, [kYCbCrFullRangeOffset+8];
$Lt_37_741122:
	.loc	20	454	0
	mov.f32 	%f342, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f343, %f178, %f342;
	fma.rn.ftz.f32 	%f159, %f343, %f339, %f159;
	fma.rn.ftz.f32 	%f161, %f343, %f340, %f161;
	fma.rn.ftz.f32 	%f163, %f343, %f341, %f163;
$Lt_37_739586:
	.loc	20	551	0
	mov.f32 	%f22, %f157;
	mov.f32 	%f24, %f159;
	mov.f32 	%f25, %f161;
	mov.f32 	%f26, %f163;
$Lt_37_731906:
$L_37_682242:
$Lt_37_731394:
	.loc	20	540	0
	@!%p130 bra 	$Lt_37_741634;
	.loc	21	268	0
	mov.f32 	%f344, %f24;
	.loc	21	269	0
	mov.f32 	%f345, %f22;
	.loc	20	558	0
	mov.f32 	%f22, %f26;
	mov.f32 	%f24, %f25;
	mov.f32 	%f25, %f344;
	mov.f32 	%f26, %f345;
$Lt_37_741634:
	@!%p8 bra 	$Lt_37_742146;
	.loc	20	57	0
	mov.u32 	%r334, 0;
	setp.ne.s32 	%p246, %r21, %r334;
	@%p246 bra 	$Lt_37_535554;
	.loc	20	59	0
	mov.f32 	%f186, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_79;
$Lt_37_535554:
	.loc	20	61	0
	mov.u32 	%r335, 64;
	setp.ne.s32 	%p247, %r21, %r335;
	@%p247 bra 	$Lt_37_535810;
	.loc	20	63	0
	mov.f32 	%f186, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_79;
$Lt_37_535810:
	.loc	20	65	0
	mov.u32 	%r336, 128;
	setp.ne.s32 	%p248, %r21, %r336;
	@%p248 bra 	$Lt_37_536066;
	.loc	20	68	0
	mov.f32 	%f186, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_79;
$Lt_37_536066:
	.loc	20	70	0
	mov.u32 	%r337, 192;
	setp.ne.s32 	%p249, %r21, %r337;
	@%p249 bra 	$Lt_37_536322;
	.loc	20	72	0
	mov.f32 	%f186, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_79;
$Lt_37_536322:
	.loc	20	76	0
	mov.f32 	%f186, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_79:
	.loc	20	564	0
	mov.f32 	%f346, 0f3f000000;   	// 0.5
	add.ftz.f32 	%f347, %f22, %f346;
	mov.f32 	%f348, 0f00000000;   	// 0
	mov.f32 	%f349, 0f00000000;   	// 0
	setp.gt.ftz.f32 	%p250, %f347, %f349;
	selp.f32 	%f350, %f347, %f348, %p250;
	min.ftz.f32 	%f22, %f350, %f186;
	mov.f32 	%f351, 0f3f000000;   	// 0.5
	add.ftz.f32 	%f352, %f24, %f351;
	mov.f32 	%f353, 0f00000000;   	// 0
	mov.f32 	%f354, 0f00000000;   	// 0
	setp.gt.ftz.f32 	%p251, %f352, %f354;
	selp.f32 	%f355, %f352, %f353, %p251;
	min.ftz.f32 	%f24, %f355, %f186;
	mov.f32 	%f356, 0f3f000000;   	// 0.5
	add.ftz.f32 	%f357, %f25, %f356;
	mov.f32 	%f358, 0f00000000;   	// 0
	mov.f32 	%f359, 0f00000000;   	// 0
	setp.gt.ftz.f32 	%p252, %f357, %f359;
	selp.f32 	%f360, %f357, %f358, %p252;
	min.ftz.f32 	%f25, %f360, %f186;
	mov.f32 	%f361, 0f3f000000;   	// 0.5
	add.ftz.f32 	%f362, %f26, %f361;
	mov.f32 	%f363, 0f00000000;   	// 0
	mov.f32 	%f364, 0f00000000;   	// 0
	setp.gt.ftz.f32 	%p253, %f362, %f364;
	selp.f32 	%f365, %f362, %f363, %p253;
	min.ftz.f32 	%f26, %f365, %f186;
$Lt_37_742146:
	st.v4.f32 	[%rd4+0], {%f22,%f24,%f25,%f26};
	.loc	22	67	0
	bra.uni 	$Lt_37_714498;
$Lt_37_714754:
	.loc	20	57	0
	@!%p3 bra 	$Lt_37_538114;
	.loc	20	59	0
	mov.f32 	%f20, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_77;
$Lt_37_538114:
	.loc	20	61	0
	mov.u32 	%r338, 64;
	setp.ne.s32 	%p254, %r14, %r338;
	@%p254 bra 	$Lt_37_538370;
	.loc	20	63	0
	mov.f32 	%f20, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_77;
$Lt_37_538370:
	.loc	20	65	0
	mov.u32 	%r339, 128;
	setp.ne.s32 	%p255, %r14, %r339;
	@%p255 bra 	$Lt_37_538626;
	.loc	20	68	0
	mov.f32 	%f20, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_77;
$Lt_37_538626:
	.loc	20	70	0
	mov.u32 	%r340, 192;
	setp.ne.s32 	%p256, %r14, %r340;
	@%p256 bra 	$Lt_37_538882;
	.loc	20	72	0
	mov.f32 	%f20, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_77;
$Lt_37_538882:
	.loc	20	76	0
	mov.f32 	%f20, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_77:
	.loc	20	469	0
	add.ftz.f32 	%f366, %f6, %f14;
	cvt.rzi.ftz.u32.f32 	%r341, %f366;
	shr.u32 	%r342, %r341, 1;
	cvt.rn.f32.u32 	%f367, %r342;
	mov.f32 	%f22, %f367;
	add.ftz.f32 	%f368, %f2, %f10;
	cvt.rzi.ftz.u32.f32 	%r343, %f368;
	shr.u32 	%r344, %r343, 1;
	cvt.rn.f32.u32 	%f369, %r344;
	mov.f32 	%f24, %f369;
	mov.f32 	%f25, %f8;
	mov.f32 	%f26, %f20;
	@!%p7 bra 	$Lt_37_743170;
	.loc	20	473	0
	mov.f32 	%f22, %f26;
	mov.f32 	%f24, %f8;
	mov.f32 	%f25, %f369;
	mov.f32 	%f26, %f367;
$Lt_37_743170:
	.loc	20	476	0
	mov.u32 	%r345, 0;
	setp.eq.s32 	%p257, %r31, %r345;
	@%p257 bra 	$Lt_37_780546;
	mov.u32 	%r346, 0;
	setp.ne.s32 	%p258, %r26, %r346;
	@%p258 bra 	$Lt_37_780546;
	.loc	20	57	0
	@!%p3 bra 	$Lt_37_539394;
	.loc	20	59	0
	mov.f32 	%f27, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_75;
$Lt_37_539394:
	.loc	20	61	0
	mov.u32 	%r347, 64;
	setp.ne.s32 	%p259, %r14, %r347;
	@%p259 bra 	$Lt_37_539650;
	.loc	20	63	0
	mov.f32 	%f27, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_75;
$Lt_37_539650:
	.loc	20	65	0
	mov.u32 	%r348, 128;
	setp.ne.s32 	%p260, %r14, %r348;
	@%p260 bra 	$Lt_37_539906;
	.loc	20	68	0
	mov.f32 	%f27, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_75;
$Lt_37_539906:
	.loc	20	70	0
	mov.u32 	%r349, 192;
	setp.ne.s32 	%p261, %r14, %r349;
	@%p261 bra 	$Lt_37_540162;
	.loc	20	72	0
	mov.f32 	%f27, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_75;
$Lt_37_540162:
	.loc	20	76	0
	mov.f32 	%f27, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_75:
	.loc	20	118	0
	and.b32 	%r37, %r13, 2048;
	mov.s32 	%r350, 0;
	setp.ne.s32 	%p15, %r37, %r350;
	@!%p15 bra 	$Lt_37_743938;
	.loc	20	100	0
	ld.const.f32 	%f370, [kYCbCrOffset+0];
	bra.uni 	$Lt_37_743682;
$Lt_37_743938:
	ld.const.f32 	%f370, [kYCbCrFullRangeOffset+0];
$Lt_37_743682:
	.loc	20	118	0
	@!%p15 bra 	$Lt_37_744450;
	.loc	20	100	0
	ld.const.f32 	%f371, [kYCbCrOffset+4];
	bra.uni 	$Lt_37_744194;
$Lt_37_744450:
	ld.const.f32 	%f371, [kYCbCrFullRangeOffset+4];
$Lt_37_744194:
	.loc	20	118	0
	@!%p15 bra 	$Lt_37_744962;
	.loc	20	100	0
	ld.const.f32 	%f372, [kYCbCrOffset+8];
	bra.uni 	$Lt_37_744706;
$Lt_37_744962:
	ld.const.f32 	%f372, [kYCbCrFullRangeOffset+8];
$Lt_37_744706:
	.loc	20	478	0
	mov.f32 	%f373, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f374, %f27, %f373;
	mul.ftz.f32 	%f375, %f374, %f370;
	sub.ftz.f32 	%f24, %f24, %f375;
	mul.ftz.f32 	%f376, %f374, %f371;
	sub.ftz.f32 	%f25, %f25, %f376;
	mul.ftz.f32 	%f377, %f374, %f372;
	sub.ftz.f32 	%f26, %f26, %f377;
$Lt_37_780546:
$Lt_37_317442:
	.loc	20	481	0
	@!%p16 bra 	$Lt_37_781058;
	mov.u32 	%r351, 0;
	setp.ne.s32 	%p262, %r39, %r351;
	@%p262 bra 	$Lt_37_781058;
	.loc	20	483	0
	mov.f32 	%f378, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p263, %f24, %f378;
	@!%p263 bra 	$Lt_37_745218;
	.loc	20	372	0
	neg.ftz.f32 	%f379, %f24;
	lg2.approx.ftz.f32 	%f380, %f379;
	mov.f32 	%f381, 0f3ee66666;   	// 0.45
	mul.ftz.f32 	%f382, %f380, %f381;
	ex2.approx.ftz.f32 	%f383, %f382;
	neg.ftz.f32 	%f42, %f383;
	bra.uni 	$LDWendi___log2f_215_71;
$Lt_37_745218:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f384, %f24;
	mov.f32 	%f385, 0f3ee66666;   	// 0.45
	mul.ftz.f32 	%f386, %f384, %f385;
	ex2.approx.ftz.f32 	%f42, %f386;
$LDWendi___log2f_215_71:
	.loc	20	483	0
	mov.f32 	%f387, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p264, %f25, %f387;
	@!%p264 bra 	$Lt_37_745730;
	.loc	20	372	0
	neg.ftz.f32 	%f388, %f25;
	lg2.approx.ftz.f32 	%f389, %f388;
	mov.f32 	%f390, 0f3ee66666;   	// 0.45
	mul.ftz.f32 	%f391, %f389, %f390;
	ex2.approx.ftz.f32 	%f392, %f391;
	neg.ftz.f32 	%f52, %f392;
	bra.uni 	$LDWendi___log2f_215_69;
$Lt_37_745730:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f393, %f25;
	mov.f32 	%f394, 0f3ee66666;   	// 0.45
	mul.ftz.f32 	%f395, %f393, %f394;
	ex2.approx.ftz.f32 	%f52, %f395;
$LDWendi___log2f_215_69:
	.loc	20	483	0
	mov.f32 	%f396, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p265, %f26, %f396;
	@!%p265 bra 	$Lt_37_746242;
	.loc	20	372	0
	neg.ftz.f32 	%f397, %f26;
	lg2.approx.ftz.f32 	%f398, %f397;
	mov.f32 	%f399, 0f3ee66666;   	// 0.45
	mul.ftz.f32 	%f400, %f398, %f399;
	ex2.approx.ftz.f32 	%f401, %f400;
	neg.ftz.f32 	%f62, %f401;
	bra.uni 	$LDWendi___log2f_215_67;
$Lt_37_746242:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f402, %f26;
	mov.f32 	%f403, 0f3ee66666;   	// 0.45
	mul.ftz.f32 	%f404, %f402, %f403;
	ex2.approx.ftz.f32 	%f62, %f404;
$LDWendi___log2f_215_67:
	.loc	20	483	0
	mov.f32 	%f24, %f42;
	mov.f32 	%f25, %f52;
	mov.f32 	%f26, %f62;
$Lt_37_781058:
$Lt_37_321282:
	.loc	20	486	0
	mov.u32 	%r352, 0;
	setp.ne.s32 	%p266, %r49, %r352;
	@%p266 bra 	$Lt_37_371970;
	setp.eq.s32 	%p267, %r26, %r31;
	@%p267 bra 	$Lt_37_372226;
$Lt_37_371970:
	.loc	20	490	0
	mov.u32 	%r353, 0;
	setp.ne.s32 	%p268, %r45, %r353;
	@%p268 bra 	$Lt_37_747010;
	mov.s32 	%r354, 256;
	setp.eq.s32 	%p26, %r21, %r354;
	mov.u32 	%r355, 256;
	setp.ne.s32 	%p269, %r14, %r355;
	@%p269 bra 	$Lt_37_747522;
	.loc	20	137	0
	mov.s32 	%r356, 512;
	setp.eq.s32 	%p28, %r44, %r356;
	@!%p26 bra 	$Lt_37_543490;
	.loc	20	139	0
	@!%p28 bra 	$Lt_37_543746;
	.loc	20	141	0
	mov.s64 	%rd6, %rd5;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_65;
$Lt_37_543746:
	.loc	20	145	0
	mov.s64 	%rd6, %rd7;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_65;
$Lt_37_543490:
	.loc	20	150	0
	@!%p28 bra 	$Lt_37_544002;
	.loc	20	152	0
	mov.s64 	%rd6, %rd8;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_65;
$Lt_37_544002:
	.loc	20	154	0
	and.b32 	%r357, %r4, 2048;
	mov.u32 	%r358, 0;
	setp.ne.s32 	%p270, %r357, %r358;
	@%p270 bra 	$Lt_37_544258;
	.loc	20	156	0
	mov.s64 	%rd6, %rd9;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_65;
$Lt_37_544258:
	.loc	20	160	0
	mov.s64 	%rd6, %rd10;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_65;
$Lt_37_747522:
	@!%p26 bra 	$Lt_37_748034;
	bra.uni 	$Lt_37_746754;
$Lt_37_748034:
	.loc	20	179	0
	mov.u32 	%r359, 512;
	setp.ne.s32 	%p271, %r44, %r359;
	@%p271 bra 	$Lt_37_545026;
	.loc	20	181	0
	mov.s64 	%rd6, %rd11;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_65;
$Lt_37_545026:
	.loc	20	183	0
	and.b32 	%r360, %r4, 2048;
	mov.u32 	%r361, 0;
	setp.ne.s32 	%p272, %r360, %r361;
	@%p272 bra 	$Lt_37_545282;
	.loc	20	185	0
	mov.s64 	%rd6, %rd12;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_65;
$Lt_37_545282:
	.loc	20	189	0
	mov.s64 	%rd6, %rd13;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_65;
$Lt_37_747010:
	mov.s32 	%r362, 0;
	setp.eq.s32 	%p32, %r43, %r362;
	mov.u32 	%r363, 512;
	setp.ne.s32 	%p273, %r46, %r363;
	@%p273 bra 	$Lt_37_748546;
	mov.s32 	%r364, 256;
	setp.eq.s32 	%p34, %r14, %r364;
	@!%p32 bra 	$Lt_37_749058;
	mov.s32 	%r365, 256;
	setp.eq.s32 	%p26, %r21, %r365;
	@!%p34 bra 	$Lt_37_749570;
	@!%p26 bra 	$Lt_37_746754;
	.loc	20	202	0
	mov.s64 	%rd6, %rd14;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_65;
$Lt_37_749570:
	.loc	20	211	0
	@!%p26 bra 	$Lt_37_546562;
	.loc	20	213	0
	mov.s64 	%rd6, %rd15;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_65;
$Lt_37_546562:
	.loc	20	217	0
	mov.s64 	%rd6, %rd16;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_65;
$Lt_37_749058:
	@!%p34 bra 	$Lt_37_750594;
	bra.uni 	$Lt_37_746754;
$Lt_37_750594:
	mov.s32 	%r366, 256;
	set.eq.u32.s32 	%r367, %r21, %r366;
	neg.s32 	%r368, %r367;
	and.b32 	%r369, %r4, 2048;
	mov.s32 	%r370, 0;
	set.eq.u32.s32 	%r371, %r369, %r370;
	neg.s32 	%r372, %r371;
	or.b32 	%r373, %r368, %r372;
	mov.u32 	%r374, 0;
	setp.eq.s32 	%p274, %r373, %r374;
	@%p274 bra 	$Lt_37_751106;
	bra.uni 	$Lt_37_746754;
$Lt_37_751106:
	.loc	20	250	0
	mov.s64 	%rd6, %rd17;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_65;
$Lt_37_748546:
	and.b32 	%r37, %r13, 2048;
	mov.s32 	%r375, 0;
	setp.eq.s32 	%p275, %r37, %r375;
	mov.pred 	%p37, %p275;
	mov.pred 	%p276, %p39;
	@!%p275 bra 	$Lt_37_751618;
	@!%p32 bra 	$Lt_37_746754;
	.loc	20	259	0
	mov.u32 	%r376, 256;
	setp.ne.s32 	%p277, %r21, %r376;
	@%p277 bra 	$Lt_37_548098;
	.loc	20	261	0
	mov.s64 	%rd6, %rd18;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_65;
$Lt_37_548098:
	.loc	20	265	0
	mov.s64 	%rd6, %rd19;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_65;
$Lt_37_751618:
	mov.s32 	%r377, 256;
	setp.eq.s32 	%p34, %r14, %r377;
	@!%p32 bra 	$Lt_37_752642;
	mov.s32 	%r378, 256;
	setp.eq.s32 	%p26, %r21, %r378;
	@!%p34 bra 	$Lt_37_753154;
	@!%p26 bra 	$Lt_37_746754;
	.loc	20	302	0
	mov.s64 	%rd6, %rd20;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_65;
$Lt_37_753154:
	.loc	20	311	0
	@!%p26 bra 	$Lt_37_549890;
	.loc	20	313	0
	mov.s64 	%rd6, %rd21;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_65;
$Lt_37_549890:
	.loc	20	317	0
	mov.s64 	%rd6, %rd22;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_65;
$Lt_37_752642:
	@!%p34 bra 	$Lt_37_754178;
	bra.uni 	$Lt_37_746754;
$Lt_37_754178:
	selp.s32 	%r379, 1, 0, %p37;
	mov.s32 	%r380, 256;
	set.eq.u32.s32 	%r381, %r21, %r380;
	neg.s32 	%r382, %r381;
	or.b32 	%r383, %r379, %r382;
	mov.u32 	%r384, 0;
	setp.eq.s32 	%p278, %r383, %r384;
	@%p278 bra 	$Lt_37_754690;
	bra.uni 	$Lt_37_746754;
$Lt_37_754690:
	.loc	20	350	0
	mov.s64 	%rd6, %rd23;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_65;
$Lt_37_746754:
	.loc	20	355	0
	mov.u64 	%rd6, 0;
$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__215_65:
	.loc	20	490	0
	ld.f32 	%f405, [%rd6+16];
	mul.ftz.f32 	%f406, %f405, %f25;
	ld.f32 	%f407, [%rd6+12];
	fma.rn.ftz.f32 	%f408, %f407, %f24, %f406;
	ld.f32 	%f409, [%rd6+20];
	fma.rn.ftz.f32 	%f410, %f409, %f26, %f408;
	ld.f32 	%f411, [%rd6+28];
	mul.ftz.f32 	%f412, %f411, %f25;
	ld.f32 	%f413, [%rd6+24];
	fma.rn.ftz.f32 	%f414, %f413, %f24, %f412;
	ld.f32 	%f415, [%rd6+32];
	fma.rn.ftz.f32 	%f416, %f415, %f26, %f414;
	ld.f32 	%f417, [%rd6+4];
	mul.ftz.f32 	%f418, %f417, %f25;
	ld.f32 	%f419, [%rd6+0];
	fma.rn.ftz.f32 	%f420, %f419, %f24, %f418;
	ld.f32 	%f421, [%rd6+8];
	fma.rn.ftz.f32 	%f24, %f421, %f26, %f420;
	mov.f32 	%f25, %f410;
	mov.f32 	%f26, %f416;
	@!%p42 bra 	$Lt_37_755458;
	.loc	20	494	0
	mov.s32 	%r385, 256;
	setp.eq.s32 	%p34, %r14, %r385;
	@!%p34 bra 	$L_37_684290;
	mov.s32 	%r386, 0;
	setp.eq.s32 	%p43, %r21, %r386;
	@%p43 bra 	$L_37_684034;
$L_37_684290:
	@!%p3 bra 	$Lt_37_782338;
	mov.u32 	%r387, 256;
	setp.ne.s32 	%p279, %r21, %r387;
	@%p279 bra 	$Lt_37_782338;
	mov.s32 	%r388, 0;
	setp.eq.s32 	%p43, %r21, %r388;
$L_37_684034:
	.loc	20	57	0
	@!%p43 bra 	$Lt_37_551426;
	.loc	20	59	0
	mov.f32 	%f83, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_63;
$Lt_37_551426:
	.loc	20	61	0
	mov.u32 	%r389, 64;
	setp.ne.s32 	%p280, %r21, %r389;
	@%p280 bra 	$Lt_37_551682;
	.loc	20	63	0
	mov.f32 	%f83, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_63;
$Lt_37_551682:
	.loc	20	65	0
	mov.u32 	%r390, 128;
	setp.ne.s32 	%p281, %r21, %r390;
	@%p281 bra 	$Lt_37_551938;
	.loc	20	68	0
	mov.f32 	%f83, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_63;
$Lt_37_551938:
	.loc	20	70	0
	mov.u32 	%r391, 192;
	setp.ne.s32 	%p282, %r21, %r391;
	@%p282 bra 	$Lt_37_552194;
	.loc	20	72	0
	mov.f32 	%f83, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_63;
$Lt_37_552194:
	.loc	20	76	0
	mov.f32 	%f83, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_63:
	.loc	20	57	0
	@!%p3 bra 	$Lt_37_552450;
	.loc	20	59	0
	mov.f32 	%f84, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_61;
$Lt_37_552450:
	.loc	20	61	0
	mov.u32 	%r392, 64;
	setp.ne.s32 	%p283, %r14, %r392;
	@%p283 bra 	$Lt_37_552706;
	.loc	20	63	0
	mov.f32 	%f84, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_61;
$Lt_37_552706:
	.loc	20	65	0
	mov.u32 	%r393, 128;
	setp.ne.s32 	%p284, %r14, %r393;
	@%p284 bra 	$Lt_37_552962;
	.loc	20	68	0
	mov.f32 	%f84, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_61;
$Lt_37_552962:
	.loc	20	70	0
	mov.u32 	%r394, 192;
	setp.ne.s32 	%p285, %r14, %r394;
	@%p285 bra 	$Lt_37_553218;
	.loc	20	72	0
	mov.f32 	%f84, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_61;
$Lt_37_553218:
	.loc	20	76	0
	mov.f32 	%f84, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_61:
	.loc	20	498	0
	div.approx.ftz.f32 	%f422, %f83, %f84;
	mul.ftz.f32 	%f22, %f22, %f422;
	bra.uni 	$Lt_37_755458;
$Lt_37_782338:
$L_37_683778:
	.loc	20	500	0
	@!%p34 bra 	$L_37_685314;
	@%p8 bra 	$L_37_685058;
$L_37_685314:
	@!%p9 bra 	$Lt_37_783362;
	mov.u32 	%r395, 256;
	setp.ne.s32 	%p286, %r21, %r395;
	@%p286 bra 	$Lt_37_783362;
$L_37_685058:
	.loc	20	57	0
	mov.u32 	%r396, 0;
	setp.ne.s32 	%p287, %r21, %r396;
	@%p287 bra 	$Lt_37_553730;
	.loc	20	59	0
	mov.f32 	%f83, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_59;
$Lt_37_553730:
	.loc	20	61	0
	mov.u32 	%r397, 64;
	setp.ne.s32 	%p288, %r21, %r397;
	@%p288 bra 	$Lt_37_553986;
	.loc	20	63	0
	mov.f32 	%f83, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_59;
$Lt_37_553986:
	.loc	20	65	0
	mov.u32 	%r398, 128;
	setp.ne.s32 	%p289, %r21, %r398;
	@%p289 bra 	$Lt_37_554242;
	.loc	20	68	0
	mov.f32 	%f83, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_59;
$Lt_37_554242:
	.loc	20	70	0
	mov.u32 	%r399, 192;
	setp.ne.s32 	%p290, %r21, %r399;
	@%p290 bra 	$Lt_37_554498;
	.loc	20	72	0
	mov.f32 	%f83, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_59;
$Lt_37_554498:
	.loc	20	76	0
	mov.f32 	%f83, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_59:
	.loc	20	57	0
	@!%p3 bra 	$Lt_37_554754;
	.loc	20	59	0
	mov.f32 	%f84, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_57;
$Lt_37_554754:
	.loc	20	61	0
	mov.u32 	%r400, 64;
	setp.ne.s32 	%p291, %r14, %r400;
	@%p291 bra 	$Lt_37_555010;
	.loc	20	63	0
	mov.f32 	%f84, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_57;
$Lt_37_555010:
	.loc	20	65	0
	mov.u32 	%r401, 128;
	setp.ne.s32 	%p292, %r14, %r401;
	@%p292 bra 	$Lt_37_555266;
	.loc	20	68	0
	mov.f32 	%f84, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_57;
$Lt_37_555266:
	.loc	20	70	0
	mov.u32 	%r402, 192;
	setp.ne.s32 	%p293, %r14, %r402;
	@%p293 bra 	$Lt_37_555522;
	.loc	20	72	0
	mov.f32 	%f84, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_57;
$Lt_37_555522:
	.loc	20	76	0
	mov.f32 	%f84, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_57:
	.loc	20	504	0
	div.approx.ftz.f32 	%f423, %f83, %f84;
	mul.ftz.f32 	%f22, %f22, %f423;
	.loc	20	57	0
	@!%p3 bra 	$Lt_37_556802;
	.loc	20	59	0
	mov.f32 	%f84, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_53;
$Lt_37_556802:
	.loc	20	61	0
	mov.u32 	%r403, 64;
	setp.ne.s32 	%p294, %r14, %r403;
	@%p294 bra 	$Lt_37_557058;
	.loc	20	63	0
	mov.f32 	%f84, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_53;
$Lt_37_557058:
	.loc	20	65	0
	mov.u32 	%r404, 128;
	setp.ne.s32 	%p295, %r14, %r404;
	@%p295 bra 	$Lt_37_557314;
	.loc	20	68	0
	mov.f32 	%f84, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_53;
$Lt_37_557314:
	.loc	20	70	0
	mov.u32 	%r405, 192;
	setp.ne.s32 	%p296, %r14, %r405;
	@%p296 bra 	$Lt_37_557570;
	.loc	20	72	0
	mov.f32 	%f84, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_53;
$Lt_37_557570:
	.loc	20	76	0
	mov.f32 	%f84, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_53:
	.loc	20	505	0
	mov.f32 	%f424, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f425, %f424, %f84;
	mul.ftz.f32 	%f24, %f24, %f425;
	.loc	20	57	0
	@!%p3 bra 	$Lt_37_558850;
	.loc	20	59	0
	mov.f32 	%f84, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_49;
$Lt_37_558850:
	.loc	20	61	0
	mov.u32 	%r406, 64;
	setp.ne.s32 	%p297, %r14, %r406;
	@%p297 bra 	$Lt_37_559106;
	.loc	20	63	0
	mov.f32 	%f84, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_49;
$Lt_37_559106:
	.loc	20	65	0
	mov.u32 	%r407, 128;
	setp.ne.s32 	%p298, %r14, %r407;
	@%p298 bra 	$Lt_37_559362;
	.loc	20	68	0
	mov.f32 	%f84, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_49;
$Lt_37_559362:
	.loc	20	70	0
	mov.u32 	%r408, 192;
	setp.ne.s32 	%p299, %r14, %r408;
	@%p299 bra 	$Lt_37_559618;
	.loc	20	72	0
	mov.f32 	%f84, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_49;
$Lt_37_559618:
	.loc	20	76	0
	mov.f32 	%f84, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_49:
	.loc	20	506	0
	mov.f32 	%f426, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f427, %f426, %f84;
	mul.ftz.f32 	%f25, %f410, %f427;
	.loc	20	57	0
	@!%p3 bra 	$Lt_37_560898;
	.loc	20	59	0
	mov.f32 	%f84, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_45;
$Lt_37_560898:
	.loc	20	61	0
	mov.u32 	%r409, 64;
	setp.ne.s32 	%p300, %r14, %r409;
	@%p300 bra 	$Lt_37_561154;
	.loc	20	63	0
	mov.f32 	%f84, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_45;
$Lt_37_561154:
	.loc	20	65	0
	mov.u32 	%r410, 128;
	setp.ne.s32 	%p301, %r14, %r410;
	@%p301 bra 	$Lt_37_561410;
	.loc	20	68	0
	mov.f32 	%f84, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_45;
$Lt_37_561410:
	.loc	20	70	0
	mov.u32 	%r411, 192;
	setp.ne.s32 	%p302, %r14, %r411;
	@%p302 bra 	$Lt_37_561666;
	.loc	20	72	0
	mov.f32 	%f84, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_45;
$Lt_37_561666:
	.loc	20	76	0
	mov.f32 	%f84, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_45:
	.loc	20	507	0
	mov.f32 	%f428, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f429, %f428, %f84;
	mul.ftz.f32 	%f26, %f416, %f429;
	bra.uni 	$Lt_37_755458;
$Lt_37_783362:
$L_37_684802:
	.loc	20	57	0
	@!%p3 bra 	$Lt_37_561922;
	.loc	20	59	0
	mov.f32 	%f83, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_43;
$Lt_37_561922:
	.loc	20	61	0
	mov.u32 	%r412, 64;
	setp.ne.s32 	%p303, %r14, %r412;
	@%p303 bra 	$Lt_37_562178;
	.loc	20	63	0
	mov.f32 	%f83, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_43;
$Lt_37_562178:
	.loc	20	65	0
	mov.u32 	%r413, 128;
	setp.ne.s32 	%p304, %r14, %r413;
	@%p304 bra 	$Lt_37_562434;
	.loc	20	68	0
	mov.f32 	%f83, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_43;
$Lt_37_562434:
	.loc	20	70	0
	mov.u32 	%r414, 192;
	setp.ne.s32 	%p305, %r14, %r414;
	@%p305 bra 	$Lt_37_562690;
	.loc	20	72	0
	mov.f32 	%f83, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_43;
$Lt_37_562690:
	.loc	20	76	0
	mov.f32 	%f83, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_43:
	.loc	20	511	0
	mov.f32 	%f430, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f431, %f83, %f430;
	mul.ftz.f32 	%f22, %f431, %f22;
	mul.ftz.f32 	%f24, %f431, %f24;
	mul.ftz.f32 	%f25, %f431, %f410;
	mul.ftz.f32 	%f26, %f431, %f416;
	bra.uni 	$Lt_37_755458;
$Lt_37_372226:
	.loc	20	486	0
	@!%p42 bra 	$Lt_37_755458;
	.loc	20	57	0
	mov.u32 	%r415, 0;
	setp.ne.s32 	%p306, %r21, %r415;
	@%p306 bra 	$Lt_37_564226;
	.loc	20	59	0
	mov.f32 	%f83, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_39;
$Lt_37_564226:
	.loc	20	61	0
	mov.u32 	%r416, 64;
	setp.ne.s32 	%p307, %r21, %r416;
	@%p307 bra 	$Lt_37_564482;
	.loc	20	63	0
	mov.f32 	%f83, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_39;
$Lt_37_564482:
	.loc	20	65	0
	mov.u32 	%r417, 128;
	setp.ne.s32 	%p308, %r21, %r417;
	@%p308 bra 	$Lt_37_564738;
	.loc	20	68	0
	mov.f32 	%f83, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_39;
$Lt_37_564738:
	.loc	20	70	0
	mov.u32 	%r418, 192;
	setp.ne.s32 	%p309, %r21, %r418;
	@%p309 bra 	$Lt_37_564994;
	.loc	20	72	0
	mov.f32 	%f83, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_39;
$Lt_37_564994:
	.loc	20	76	0
	mov.f32 	%f83, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_39:
	.loc	20	57	0
	@!%p3 bra 	$Lt_37_565250;
	.loc	20	59	0
	mov.f32 	%f84, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_37;
$Lt_37_565250:
	.loc	20	61	0
	mov.u32 	%r419, 64;
	setp.ne.s32 	%p310, %r14, %r419;
	@%p310 bra 	$Lt_37_565506;
	.loc	20	63	0
	mov.f32 	%f84, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_37;
$Lt_37_565506:
	.loc	20	65	0
	mov.u32 	%r420, 128;
	setp.ne.s32 	%p311, %r14, %r420;
	@%p311 bra 	$Lt_37_565762;
	.loc	20	68	0
	mov.f32 	%f84, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_37;
$Lt_37_565762:
	.loc	20	70	0
	mov.u32 	%r421, 192;
	setp.ne.s32 	%p312, %r14, %r421;
	@%p312 bra 	$Lt_37_566018;
	.loc	20	72	0
	mov.f32 	%f84, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_37;
$Lt_37_566018:
	.loc	20	76	0
	mov.f32 	%f84, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_37:
	.loc	20	517	0
	div.approx.ftz.f32 	%f432, %f83, %f84;
	mul.ftz.f32 	%f22, %f432, %f22;
	mul.ftz.f32 	%f24, %f432, %f24;
	mul.ftz.f32 	%f25, %f432, %f25;
	mul.ftz.f32 	%f26, %f432, %f26;
$Lt_37_755458:
$Lt_37_374530:
	.loc	20	520	0
	@!%p78 bra 	$Lt_37_783874;
	mov.u32 	%r422, 0;
	setp.ne.s32 	%p313, %r40, %r422;
	@%p313 bra 	$Lt_37_783874;
	.loc	20	522	0
	mov.f32 	%f433, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p314, %f24, %f433;
	@!%p314 bra 	$Lt_37_755970;
	.loc	20	372	0
	neg.ftz.f32 	%f434, %f24;
	lg2.approx.ftz.f32 	%f435, %f434;
	mov.f32 	%f436, 0f400e38e4;   	// 2.22222
	mul.ftz.f32 	%f437, %f435, %f436;
	ex2.approx.ftz.f32 	%f438, %f437;
	neg.ftz.f32 	%f102, %f438;
	bra.uni 	$LDWendi___log2f_215_35;
$Lt_37_755970:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f439, %f24;
	mov.f32 	%f440, 0f400e38e4;   	// 2.22222
	mul.ftz.f32 	%f441, %f439, %f440;
	ex2.approx.ftz.f32 	%f102, %f441;
$LDWendi___log2f_215_35:
	.loc	20	522	0
	mov.f32 	%f442, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p315, %f25, %f442;
	@!%p315 bra 	$Lt_37_756482;
	.loc	20	372	0
	neg.ftz.f32 	%f443, %f25;
	lg2.approx.ftz.f32 	%f444, %f443;
	mov.f32 	%f445, 0f400e38e4;   	// 2.22222
	mul.ftz.f32 	%f446, %f444, %f445;
	ex2.approx.ftz.f32 	%f447, %f446;
	neg.ftz.f32 	%f112, %f447;
	bra.uni 	$LDWendi___log2f_215_33;
$Lt_37_756482:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f448, %f25;
	mov.f32 	%f449, 0f400e38e4;   	// 2.22222
	mul.ftz.f32 	%f450, %f448, %f449;
	ex2.approx.ftz.f32 	%f112, %f450;
$LDWendi___log2f_215_33:
	.loc	20	522	0
	mov.f32 	%f451, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p316, %f26, %f451;
	@!%p316 bra 	$Lt_37_756994;
	.loc	20	372	0
	neg.ftz.f32 	%f452, %f26;
	lg2.approx.ftz.f32 	%f453, %f452;
	mov.f32 	%f454, 0f400e38e4;   	// 2.22222
	mul.ftz.f32 	%f455, %f453, %f454;
	ex2.approx.ftz.f32 	%f456, %f455;
	neg.ftz.f32 	%f122, %f456;
	bra.uni 	$LDWendi___log2f_215_31;
$Lt_37_756994:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f457, %f26;
	mov.f32 	%f458, 0f400e38e4;   	// 2.22222
	mul.ftz.f32 	%f459, %f457, %f458;
	ex2.approx.ftz.f32 	%f122, %f459;
$LDWendi___log2f_215_31:
	.loc	20	522	0
	mov.f32 	%f24, %f102;
	mov.f32 	%f25, %f112;
	mov.f32 	%f26, %f122;
$Lt_37_783874:
$Lt_37_376578:
	.loc	20	525	0
	mov.u32 	%r423, 0;
	setp.eq.s32 	%p317, %r26, %r423;
	@%p317 bra 	$Lt_37_784386;
	mov.u32 	%r424, 0;
	setp.ne.s32 	%p318, %r31, %r424;
	@%p318 bra 	$Lt_37_784386;
	.loc	20	57	0
	mov.u32 	%r425, 0;
	setp.ne.s32 	%p319, %r21, %r425;
	@%p319 bra 	$Lt_37_567042;
	.loc	20	59	0
	mov.f32 	%f126, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_29;
$Lt_37_567042:
	.loc	20	61	0
	mov.u32 	%r426, 64;
	setp.ne.s32 	%p320, %r21, %r426;
	@%p320 bra 	$Lt_37_567298;
	.loc	20	63	0
	mov.f32 	%f126, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_29;
$Lt_37_567298:
	.loc	20	65	0
	mov.u32 	%r427, 128;
	setp.ne.s32 	%p321, %r21, %r427;
	@%p321 bra 	$Lt_37_567554;
	.loc	20	68	0
	mov.f32 	%f126, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_29;
$Lt_37_567554:
	.loc	20	70	0
	mov.u32 	%r428, 192;
	setp.ne.s32 	%p322, %r21, %r428;
	@%p322 bra 	$Lt_37_567810;
	.loc	20	72	0
	mov.f32 	%f126, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_29;
$Lt_37_567810:
	.loc	20	76	0
	mov.f32 	%f126, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_29:
	.loc	20	107	0
	and.b32 	%r128, %r4, 2048;
	mov.s32 	%r429, 0;
	setp.ne.s32 	%p89, %r128, %r429;
	@!%p89 bra 	$Lt_37_757762;
	.loc	20	100	0
	ld.const.f32 	%f460, [kYCbCrOffset+0];
	bra.uni 	$Lt_37_757506;
$Lt_37_757762:
	ld.const.f32 	%f460, [kYCbCrFullRangeOffset+0];
$Lt_37_757506:
	.loc	20	107	0
	@!%p89 bra 	$Lt_37_758274;
	.loc	20	100	0
	ld.const.f32 	%f461, [kYCbCrOffset+4];
	bra.uni 	$Lt_37_758018;
$Lt_37_758274:
	ld.const.f32 	%f461, [kYCbCrFullRangeOffset+4];
$Lt_37_758018:
	.loc	20	107	0
	@!%p89 bra 	$Lt_37_758786;
	.loc	20	100	0
	ld.const.f32 	%f462, [kYCbCrOffset+8];
	bra.uni 	$Lt_37_758530;
$Lt_37_758786:
	ld.const.f32 	%f462, [kYCbCrFullRangeOffset+8];
$Lt_37_758530:
	.loc	20	527	0
	mov.f32 	%f463, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f464, %f126, %f463;
	fma.rn.ftz.f32 	%f24, %f464, %f460, %f24;
	fma.rn.ftz.f32 	%f25, %f464, %f461, %f25;
	fma.rn.ftz.f32 	%f26, %f464, %f462, %f26;
$Lt_37_784386:
$Lt_37_382978:
	.loc	20	525	0
	@!%p90 bra 	$Lt_37_759554;
	.loc	20	532	0
	mov.u32 	%r430, 8;
	setp.ne.s32 	%p323, %r131, %r430;
	@%p323 bra 	$L_37_686594;
	mov.u32 	%r431, 12;
	setp.eq.s32 	%p324, %r130, %r431;
	@%p324 bra 	$Lt_37_785154;
$L_37_686594:
	mov.u32 	%r432, 12;
	setp.eq.s32 	%p325, %r131, %r432;
	@%p325 bra 	$Lt_37_785154;
	mov.u32 	%r433, 0;
	setp.ne.s32 	%p326, %r131, %r433;
	@%p326 bra 	$L_37_685826;
$Lt_37_785154:
$L_37_686082:
	.loc	20	57	0
	mov.u32 	%r434, 0;
	setp.ne.s32 	%p327, %r21, %r434;
	@%p327 bra 	$Lt_37_570370;
	.loc	20	59	0
	mov.f32 	%f132, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_25;
$Lt_37_570370:
	.loc	20	61	0
	mov.u32 	%r435, 64;
	setp.ne.s32 	%p328, %r21, %r435;
	@%p328 bra 	$Lt_37_570626;
	.loc	20	63	0
	mov.f32 	%f132, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_25;
$Lt_37_570626:
	.loc	20	65	0
	mov.u32 	%r436, 128;
	setp.ne.s32 	%p329, %r21, %r436;
	@%p329 bra 	$Lt_37_570882;
	.loc	20	68	0
	mov.f32 	%f132, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_25;
$Lt_37_570882:
	.loc	20	70	0
	mov.u32 	%r437, 192;
	setp.ne.s32 	%p330, %r21, %r437;
	@%p330 bra 	$Lt_37_571138;
	.loc	20	72	0
	mov.f32 	%f132, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_25;
$Lt_37_571138:
	.loc	20	76	0
	mov.f32 	%f132, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_25:
	.loc	20	536	0
	mov.f32 	%f22, %f132;
	bra.uni 	$Lt_37_759554;
$L_37_685826:
	.loc	20	540	0
	mov.s32 	%r438, 12;
	setp.eq.s32 	%p99, %r130, %r438;
	mov.s32 	%r439, 4;
	set.eq.u32.s32 	%r440, %r131, %r439;
	neg.s32 	%r441, %r440;
	selp.s32 	%r442, 1, 0, %p99;
	mov.s32 	%r443, 8;
	set.eq.u32.s32 	%r444, %r130, %r443;
	neg.s32 	%r445, %r444;
	or.b32 	%r446, %r442, %r445;
	and.b32 	%r447, %r441, %r446;
	mov.u32 	%r448, 0;
	setp.eq.s32 	%p331, %r447, %r448;
	@%p331 bra 	$Lt_37_759810;
	.loc	20	410	0
	mov.f32 	%f465, %f24;
	mov.f32 	%f134, %f465;
	mov.f32 	%f466, %f25;
	mov.f32 	%f136, %f466;
	mov.f32 	%f467, %f26;
	mov.f32 	%f138, %f467;
	.loc	20	57	0
	mov.s32 	%r449, 0;
	setp.eq.s32 	%p43, %r21, %r449;
	@!%p43 bra 	$Lt_37_571650;
	.loc	20	59	0
	mov.f32 	%f139, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_23;
$Lt_37_571650:
	.loc	20	61	0
	mov.u32 	%r450, 64;
	setp.ne.s32 	%p332, %r21, %r450;
	@%p332 bra 	$Lt_37_571906;
	.loc	20	63	0
	mov.f32 	%f139, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_23;
$Lt_37_571906:
	.loc	20	65	0
	mov.u32 	%r451, 128;
	setp.ne.s32 	%p333, %r21, %r451;
	@%p333 bra 	$Lt_37_572162;
	.loc	20	68	0
	mov.f32 	%f139, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_23;
$Lt_37_572162:
	.loc	20	70	0
	mov.u32 	%r452, 192;
	setp.ne.s32 	%p334, %r21, %r452;
	@%p334 bra 	$Lt_37_572418;
	.loc	20	72	0
	mov.f32 	%f139, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_23;
$Lt_37_572418:
	.loc	20	76	0
	mov.f32 	%f139, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_23:
	.loc	20	413	0
	mov.u32 	%r453, 0;
	setp.eq.s32 	%p335, %r26, %r453;
	@%p335 bra 	$Lt_37_760066;
	.loc	20	57	0
	@!%p43 bra 	$Lt_37_572930;
	.loc	20	59	0
	mov.f32 	%f140, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_21;
$Lt_37_572930:
	.loc	20	61	0
	mov.u32 	%r454, 64;
	setp.ne.s32 	%p336, %r21, %r454;
	@%p336 bra 	$Lt_37_573186;
	.loc	20	63	0
	mov.f32 	%f140, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_21;
$Lt_37_573186:
	.loc	20	65	0
	mov.u32 	%r455, 128;
	setp.ne.s32 	%p337, %r21, %r455;
	@%p337 bra 	$Lt_37_573442;
	.loc	20	68	0
	mov.f32 	%f140, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_21;
$Lt_37_573442:
	.loc	20	70	0
	mov.u32 	%r456, 192;
	setp.ne.s32 	%p338, %r21, %r456;
	@%p338 bra 	$Lt_37_573698;
	.loc	20	72	0
	mov.f32 	%f140, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_21;
$Lt_37_573698:
	.loc	20	76	0
	mov.f32 	%f140, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_21:
	.loc	20	118	0
	and.b32 	%r128, %r4, 2048;
	mov.s32 	%r457, 0;
	setp.ne.s32 	%p89, %r128, %r457;
	@!%p89 bra 	$Lt_37_760834;
	.loc	20	100	0
	ld.const.f32 	%f468, [kYCbCrOffset+0];
	bra.uni 	$Lt_37_760578;
$Lt_37_760834:
	ld.const.f32 	%f468, [kYCbCrFullRangeOffset+0];
$Lt_37_760578:
	.loc	20	118	0
	@!%p89 bra 	$Lt_37_761346;
	.loc	20	100	0
	ld.const.f32 	%f469, [kYCbCrOffset+4];
	bra.uni 	$Lt_37_761090;
$Lt_37_761346:
	ld.const.f32 	%f469, [kYCbCrFullRangeOffset+4];
$Lt_37_761090:
	.loc	20	118	0
	@!%p89 bra 	$Lt_37_761858;
	.loc	20	100	0
	ld.const.f32 	%f470, [kYCbCrOffset+8];
	bra.uni 	$Lt_37_761602;
$Lt_37_761858:
	ld.const.f32 	%f470, [kYCbCrFullRangeOffset+8];
$Lt_37_761602:
	.loc	20	415	0
	mov.f32 	%f471, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f472, %f140, %f471;
	mul.ftz.f32 	%f473, %f472, %f468;
	sub.ftz.f32 	%f134, %f465, %f473;
	mul.ftz.f32 	%f474, %f472, %f469;
	sub.ftz.f32 	%f136, %f466, %f474;
	mul.ftz.f32 	%f475, %f472, %f470;
	sub.ftz.f32 	%f138, %f467, %f475;
$Lt_37_760066:
	.loc	20	418	0
	rcp.approx.ftz.f32 	%f476, %f139;
	mul.ftz.f32 	%f477, %f476, %f22;
	mul.ftz.f32 	%f134, %f477, %f134;
	.loc	20	419	0
	mul.ftz.f32 	%f136, %f477, %f136;
	.loc	20	420	0
	mul.ftz.f32 	%f138, %f477, %f138;
	.loc	20	422	0
	mov.u32 	%r458, 0;
	setp.eq.s32 	%p339, %r26, %r458;
	@%p339 bra 	$Lt_37_762114;
	.loc	20	57	0
	@!%p43 bra 	$Lt_37_576002;
	.loc	20	59	0
	mov.f32 	%f151, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_17;
$Lt_37_576002:
	.loc	20	61	0
	mov.u32 	%r459, 64;
	setp.ne.s32 	%p340, %r21, %r459;
	@%p340 bra 	$Lt_37_576258;
	.loc	20	63	0
	mov.f32 	%f151, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_17;
$Lt_37_576258:
	.loc	20	65	0
	mov.u32 	%r460, 128;
	setp.ne.s32 	%p341, %r21, %r460;
	@%p341 bra 	$Lt_37_576514;
	.loc	20	68	0
	mov.f32 	%f151, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_17;
$Lt_37_576514:
	.loc	20	70	0
	mov.u32 	%r461, 192;
	setp.ne.s32 	%p342, %r21, %r461;
	@%p342 bra 	$Lt_37_576770;
	.loc	20	72	0
	mov.f32 	%f151, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_17;
$Lt_37_576770:
	.loc	20	76	0
	mov.f32 	%f151, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_17:
	.loc	20	107	0
	and.b32 	%r128, %r4, 2048;
	mov.s32 	%r462, 0;
	setp.ne.s32 	%p89, %r128, %r462;
	@!%p89 bra 	$Lt_37_762882;
	.loc	20	100	0
	ld.const.f32 	%f478, [kYCbCrOffset+0];
	bra.uni 	$Lt_37_762626;
$Lt_37_762882:
	ld.const.f32 	%f478, [kYCbCrFullRangeOffset+0];
$Lt_37_762626:
	.loc	20	107	0
	@!%p89 bra 	$Lt_37_763394;
	.loc	20	100	0
	ld.const.f32 	%f479, [kYCbCrOffset+4];
	bra.uni 	$Lt_37_763138;
$Lt_37_763394:
	ld.const.f32 	%f479, [kYCbCrFullRangeOffset+4];
$Lt_37_763138:
	.loc	20	107	0
	@!%p89 bra 	$Lt_37_763906;
	.loc	20	100	0
	ld.const.f32 	%f480, [kYCbCrOffset+8];
	bra.uni 	$Lt_37_763650;
$Lt_37_763906:
	ld.const.f32 	%f480, [kYCbCrFullRangeOffset+8];
$Lt_37_763650:
	.loc	20	424	0
	mov.f32 	%f481, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f482, %f151, %f481;
	fma.rn.ftz.f32 	%f134, %f482, %f478, %f134;
	fma.rn.ftz.f32 	%f136, %f482, %f479, %f136;
	fma.rn.ftz.f32 	%f138, %f482, %f480, %f138;
$Lt_37_762114:
	.loc	20	543	0
	mov.f32 	%f24, %f134;
	mov.f32 	%f25, %f136;
	mov.f32 	%f26, %f138;
	@!%p99 bra 	$Lt_37_759554;
	.loc	20	57	0
	@!%p43 bra 	$Lt_37_579074;
	.loc	20	59	0
	mov.f32 	%f132, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_13;
$Lt_37_579074:
	.loc	20	61	0
	mov.u32 	%r463, 64;
	setp.ne.s32 	%p343, %r21, %r463;
	@%p343 bra 	$Lt_37_579330;
	.loc	20	63	0
	mov.f32 	%f132, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_13;
$Lt_37_579330:
	.loc	20	65	0
	mov.u32 	%r464, 128;
	setp.ne.s32 	%p344, %r21, %r464;
	@%p344 bra 	$Lt_37_579586;
	.loc	20	68	0
	mov.f32 	%f132, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_13;
$Lt_37_579586:
	.loc	20	70	0
	mov.u32 	%r465, 192;
	setp.ne.s32 	%p345, %r21, %r465;
	@%p345 bra 	$Lt_37_579842;
	.loc	20	72	0
	mov.f32 	%f132, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_13;
$Lt_37_579842:
	.loc	20	76	0
	mov.f32 	%f132, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_13:
	.loc	20	546	0
	mov.f32 	%f22, %f132;
	bra.uni 	$Lt_37_759554;
$Lt_37_759810:
	.loc	20	433	0
	mov.f32 	%f157, %f22;
	mov.f32 	%f483, %f24;
	mov.f32 	%f159, %f483;
	mov.f32 	%f484, %f25;
	mov.f32 	%f161, %f484;
	mov.f32 	%f485, %f26;
	mov.f32 	%f163, %f485;
	.loc	20	435	0
	mov.u32 	%r466, 0;
	setp.eq.s32 	%p346, %r26, %r466;
	@%p346 bra 	$Lt_37_764674;
	.loc	20	57	0
	mov.u32 	%r467, 0;
	setp.ne.s32 	%p347, %r21, %r467;
	@%p347 bra 	$Lt_37_580354;
	.loc	20	59	0
	mov.f32 	%f164, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_11;
$Lt_37_580354:
	.loc	20	61	0
	mov.u32 	%r468, 64;
	setp.ne.s32 	%p348, %r21, %r468;
	@%p348 bra 	$Lt_37_580610;
	.loc	20	63	0
	mov.f32 	%f164, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_11;
$Lt_37_580610:
	.loc	20	65	0
	mov.u32 	%r469, 128;
	setp.ne.s32 	%p349, %r21, %r469;
	@%p349 bra 	$Lt_37_580866;
	.loc	20	68	0
	mov.f32 	%f164, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_11;
$Lt_37_580866:
	.loc	20	70	0
	mov.u32 	%r470, 192;
	setp.ne.s32 	%p350, %r21, %r470;
	@%p350 bra 	$Lt_37_581122;
	.loc	20	72	0
	mov.f32 	%f164, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_11;
$Lt_37_581122:
	.loc	20	76	0
	mov.f32 	%f164, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_11:
	.loc	20	118	0
	and.b32 	%r128, %r4, 2048;
	mov.s32 	%r471, 0;
	setp.ne.s32 	%p89, %r128, %r471;
	@!%p89 bra 	$Lt_37_765442;
	.loc	20	100	0
	ld.const.f32 	%f486, [kYCbCrOffset+0];
	bra.uni 	$Lt_37_765186;
$Lt_37_765442:
	ld.const.f32 	%f486, [kYCbCrFullRangeOffset+0];
$Lt_37_765186:
	.loc	20	118	0
	@!%p89 bra 	$Lt_37_765954;
	.loc	20	100	0
	ld.const.f32 	%f487, [kYCbCrOffset+4];
	bra.uni 	$Lt_37_765698;
$Lt_37_765954:
	ld.const.f32 	%f487, [kYCbCrFullRangeOffset+4];
$Lt_37_765698:
	.loc	20	118	0
	@!%p89 bra 	$Lt_37_766466;
	.loc	20	100	0
	ld.const.f32 	%f488, [kYCbCrOffset+8];
	bra.uni 	$Lt_37_766210;
$Lt_37_766466:
	ld.const.f32 	%f488, [kYCbCrFullRangeOffset+8];
$Lt_37_766210:
	.loc	20	437	0
	mov.f32 	%f489, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f490, %f164, %f489;
	mul.ftz.f32 	%f491, %f490, %f486;
	sub.ftz.f32 	%f159, %f483, %f491;
	mul.ftz.f32 	%f492, %f490, %f487;
	sub.ftz.f32 	%f161, %f484, %f492;
	mul.ftz.f32 	%f493, %f490, %f488;
	sub.ftz.f32 	%f163, %f485, %f493;
$Lt_37_764674:
	mov.f32 	%f494, 0fb70637bd;   	// -8e-006
	add.ftz.f32 	%f495, %f22, %f494;
	mov.f32 	%f496, 0f00000000;   	// 0
	setp.le.ftz.f32 	%p351, %f495, %f496;
	@!%p351 bra 	$Lt_37_766978;
	mov.f32 	%f163, 0f00000000;   	// 0
	mov.f32 	%f161, 0f00000000;   	// 0
	mov.f32 	%f159, 0f00000000;   	// 0
	mov.f32 	%f157, 0f00000000;   	// 0
	bra.uni 	$Lt_37_766722;
$Lt_37_766978:
	.loc	20	57	0
	mov.u32 	%r472, 0;
	setp.ne.s32 	%p352, %r21, %r472;
	@%p352 bra 	$Lt_37_583426;
	.loc	20	59	0
	mov.f32 	%f176, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_7;
$Lt_37_583426:
	.loc	20	61	0
	mov.u32 	%r473, 64;
	setp.ne.s32 	%p353, %r21, %r473;
	@%p353 bra 	$Lt_37_583682;
	.loc	20	63	0
	mov.f32 	%f176, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_7;
$Lt_37_583682:
	.loc	20	65	0
	mov.u32 	%r474, 128;
	setp.ne.s32 	%p354, %r21, %r474;
	@%p354 bra 	$Lt_37_583938;
	.loc	20	68	0
	mov.f32 	%f176, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_7;
$Lt_37_583938:
	.loc	20	70	0
	mov.u32 	%r475, 192;
	setp.ne.s32 	%p355, %r21, %r475;
	@%p355 bra 	$Lt_37_584194;
	.loc	20	72	0
	mov.f32 	%f176, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_7;
$Lt_37_584194:
	.loc	20	76	0
	mov.f32 	%f176, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_7:
	.loc	20	447	0
	div.approx.ftz.f32 	%f497, %f176, %f22;
	mul.ftz.f32 	%f159, %f497, %f159;
	.loc	20	448	0
	mul.ftz.f32 	%f161, %f497, %f161;
	.loc	20	449	0
	mul.ftz.f32 	%f163, %f497, %f163;
$Lt_37_766722:
	.loc	20	452	0
	mov.u32 	%r476, 0;
	setp.eq.s32 	%p356, %r26, %r476;
	@%p356 bra 	$Lt_37_767234;
	.loc	20	57	0
	mov.u32 	%r477, 0;
	setp.ne.s32 	%p357, %r21, %r477;
	@%p357 bra 	$Lt_37_584706;
	.loc	20	59	0
	mov.f32 	%f178, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_5;
$Lt_37_584706:
	.loc	20	61	0
	mov.u32 	%r478, 64;
	setp.ne.s32 	%p358, %r21, %r478;
	@%p358 bra 	$Lt_37_584962;
	.loc	20	63	0
	mov.f32 	%f178, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_5;
$Lt_37_584962:
	.loc	20	65	0
	mov.u32 	%r479, 128;
	setp.ne.s32 	%p359, %r21, %r479;
	@%p359 bra 	$Lt_37_585218;
	.loc	20	68	0
	mov.f32 	%f178, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_5;
$Lt_37_585218:
	.loc	20	70	0
	mov.u32 	%r480, 192;
	setp.ne.s32 	%p360, %r21, %r480;
	@%p360 bra 	$Lt_37_585474;
	.loc	20	72	0
	mov.f32 	%f178, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_5;
$Lt_37_585474:
	.loc	20	76	0
	mov.f32 	%f178, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_5:
	.loc	20	107	0
	and.b32 	%r128, %r4, 2048;
	mov.s32 	%r481, 0;
	setp.ne.s32 	%p89, %r128, %r481;
	@!%p89 bra 	$Lt_37_768002;
	.loc	20	100	0
	ld.const.f32 	%f498, [kYCbCrOffset+0];
	bra.uni 	$Lt_37_767746;
$Lt_37_768002:
	ld.const.f32 	%f498, [kYCbCrFullRangeOffset+0];
$Lt_37_767746:
	.loc	20	107	0
	@!%p89 bra 	$Lt_37_768514;
	.loc	20	100	0
	ld.const.f32 	%f499, [kYCbCrOffset+4];
	bra.uni 	$Lt_37_768258;
$Lt_37_768514:
	ld.const.f32 	%f499, [kYCbCrFullRangeOffset+4];
$Lt_37_768258:
	.loc	20	107	0
	@!%p89 bra 	$Lt_37_769026;
	.loc	20	100	0
	ld.const.f32 	%f500, [kYCbCrOffset+8];
	bra.uni 	$Lt_37_768770;
$Lt_37_769026:
	ld.const.f32 	%f500, [kYCbCrFullRangeOffset+8];
$Lt_37_768770:
	.loc	20	454	0
	mov.f32 	%f501, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f502, %f178, %f501;
	fma.rn.ftz.f32 	%f159, %f502, %f498, %f159;
	fma.rn.ftz.f32 	%f161, %f502, %f499, %f161;
	fma.rn.ftz.f32 	%f163, %f502, %f500, %f163;
$Lt_37_767234:
	.loc	20	551	0
	mov.f32 	%f22, %f157;
	mov.f32 	%f24, %f159;
	mov.f32 	%f25, %f161;
	mov.f32 	%f26, %f163;
$Lt_37_759554:
$L_37_685570:
$Lt_37_759042:
	.loc	20	540	0
	@!%p130 bra 	$Lt_37_769282;
	.loc	21	268	0
	mov.f32 	%f503, %f24;
	.loc	21	269	0
	mov.f32 	%f504, %f22;
	.loc	20	558	0
	mov.f32 	%f22, %f26;
	mov.f32 	%f24, %f25;
	mov.f32 	%f25, %f503;
	mov.f32 	%f26, %f504;
$Lt_37_769282:
	@!%p8 bra 	$Lt_37_769794;
	.loc	20	57	0
	mov.u32 	%r482, 0;
	setp.ne.s32 	%p361, %r21, %r482;
	@%p361 bra 	$Lt_37_588034;
	.loc	20	59	0
	mov.f32 	%f186, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_1;
$Lt_37_588034:
	.loc	20	61	0
	mov.u32 	%r483, 64;
	setp.ne.s32 	%p362, %r21, %r483;
	@%p362 bra 	$Lt_37_588290;
	.loc	20	63	0
	mov.f32 	%f186, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_1;
$Lt_37_588290:
	.loc	20	65	0
	mov.u32 	%r484, 128;
	setp.ne.s32 	%p363, %r21, %r484;
	@%p363 bra 	$Lt_37_588546;
	.loc	20	68	0
	mov.f32 	%f186, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_1;
$Lt_37_588546:
	.loc	20	70	0
	mov.u32 	%r485, 192;
	setp.ne.s32 	%p364, %r21, %r485;
	@%p364 bra 	$Lt_37_588802;
	.loc	20	72	0
	mov.f32 	%f186, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_215_1;
$Lt_37_588802:
	.loc	20	76	0
	mov.f32 	%f186, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_215_1:
	.loc	20	564	0
	mov.f32 	%f505, 0f3f000000;   	// 0.5
	add.ftz.f32 	%f506, %f22, %f505;
	mov.f32 	%f507, 0f00000000;   	// 0
	mov.f32 	%f508, 0f00000000;   	// 0
	setp.gt.ftz.f32 	%p365, %f506, %f508;
	selp.f32 	%f509, %f506, %f507, %p365;
	min.ftz.f32 	%f22, %f509, %f186;
	mov.f32 	%f510, 0f3f000000;   	// 0.5
	add.ftz.f32 	%f511, %f24, %f510;
	mov.f32 	%f512, 0f00000000;   	// 0
	mov.f32 	%f513, 0f00000000;   	// 0
	setp.gt.ftz.f32 	%p366, %f511, %f513;
	selp.f32 	%f514, %f511, %f512, %p366;
	min.ftz.f32 	%f24, %f514, %f186;
	mov.f32 	%f515, 0f3f000000;   	// 0.5
	add.ftz.f32 	%f516, %f25, %f515;
	mov.f32 	%f517, 0f00000000;   	// 0
	mov.f32 	%f518, 0f00000000;   	// 0
	setp.gt.ftz.f32 	%p367, %f516, %f518;
	selp.f32 	%f519, %f516, %f517, %p367;
	min.ftz.f32 	%f25, %f519, %f186;
	mov.f32 	%f520, 0f3f000000;   	// 0.5
	add.ftz.f32 	%f521, %f26, %f520;
	mov.f32 	%f522, 0f00000000;   	// 0
	mov.f32 	%f523, 0f00000000;   	// 0
	setp.gt.ftz.f32 	%p368, %f521, %f523;
	selp.f32 	%f524, %f521, %f522, %p368;
	min.ftz.f32 	%f26, %f524, %f186;
$Lt_37_769794:
	st.v4.f32 	[%rd4+0], {%f22,%f24,%f25,%f26};
$Lt_37_714498:
	.loc	22	87	0
	ret;
$LDWend__Z26ConvertPixelTwo_422_To_4446float4S_14IR_PixelFormatPS_S1_S0_:
	} // _Z26ConvertPixelTwo_422_To_4446float4S_14IR_PixelFormatPS_S1_S0_

	.visible .func _Z33ConvertPixelTwo_422_8u_To_444_32f6uchar4S_14IR_PixelFormatP6float4S2_S0_ (.param .align 4 .b8 __cudaparmf1__Z33ConvertPixelTwo_422_8u_To_444_32f6uchar4S_14IR_PixelFormatP6float4S2_S0_[4], .param .align 4 .b8 __cudaparmf2__Z33ConvertPixelTwo_422_8u_To_444_32f6uchar4S_14IR_PixelFormatP6float4S2_S0_[4], .param .s32 __cudaparmf3__Z33ConvertPixelTwo_422_8u_To_444_32f6uchar4S_14IR_PixelFormatP6float4S2_S0_, .param .u64 __cudaparmf4__Z33ConvertPixelTwo_422_8u_To_444_32f6uchar4S_14IR_PixelFormatP6float4S2_S0_, .param .u64 __cudaparmf5__Z33ConvertPixelTwo_422_8u_To_444_32f6uchar4S_14IR_PixelFormatP6float4S2_S0_, .param .s32 __cudaparmf6__Z33ConvertPixelTwo_422_8u_To_444_32f6uchar4S_14IR_PixelFormatP6float4S2_S0_)
	{
	.reg .u32 %r<511>;
	.reg .u64 %rd<25>;
	.reg .f32 %f<518>;
	.reg .pred %p<370>;
	.loc	22	99	0
$LDWbegin__Z33ConvertPixelTwo_422_8u_To_444_32f6uchar4S_14IR_PixelFormatP6float4S2_S0_:
	ld.param.u8 	%r1, [__cudaparmf1__Z33ConvertPixelTwo_422_8u_To_444_32f6uchar4S_14IR_PixelFormatP6float4S2_S0_+0];
	mov.s32 	%r2, %r1;
	ld.param.u8 	%r3, [__cudaparmf1__Z33ConvertPixelTwo_422_8u_To_444_32f6uchar4S_14IR_PixelFormatP6float4S2_S0_+1];
	mov.s32 	%r4, %r3;
	ld.param.u8 	%r5, [__cudaparmf1__Z33ConvertPixelTwo_422_8u_To_444_32f6uchar4S_14IR_PixelFormatP6float4S2_S0_+2];
	mov.s32 	%r6, %r5;
	ld.param.u8 	%r7, [__cudaparmf1__Z33ConvertPixelTwo_422_8u_To_444_32f6uchar4S_14IR_PixelFormatP6float4S2_S0_+3];
	mov.s32 	%r8, %r7;
	ld.param.u8 	%r9, [__cudaparmf2__Z33ConvertPixelTwo_422_8u_To_444_32f6uchar4S_14IR_PixelFormatP6float4S2_S0_+0];
	mov.s32 	%r10, %r9;
	ld.param.u8 	%r11, [__cudaparmf2__Z33ConvertPixelTwo_422_8u_To_444_32f6uchar4S_14IR_PixelFormatP6float4S2_S0_+1];
	mov.s32 	%r12, %r11;
	ld.param.u8 	%r13, [__cudaparmf2__Z33ConvertPixelTwo_422_8u_To_444_32f6uchar4S_14IR_PixelFormatP6float4S2_S0_+2];
	mov.s32 	%r14, %r13;
	ld.param.u8 	%r15, [__cudaparmf2__Z33ConvertPixelTwo_422_8u_To_444_32f6uchar4S_14IR_PixelFormatP6float4S2_S0_+3];
	mov.s32 	%r16, %r15;
	ld.param.u32 	%r17, [__cudaparmf3__Z33ConvertPixelTwo_422_8u_To_444_32f6uchar4S_14IR_PixelFormatP6float4S2_S0_];
	mov.s32 	%r18, %r17;
	ld.param.u64 	%rd1, [__cudaparmf4__Z33ConvertPixelTwo_422_8u_To_444_32f6uchar4S_14IR_PixelFormatP6float4S2_S0_];
	mov.s64 	%rd2, %rd1;
	ld.param.u64 	%rd3, [__cudaparmf5__Z33ConvertPixelTwo_422_8u_To_444_32f6uchar4S_14IR_PixelFormatP6float4S2_S0_];
	mov.s64 	%rd4, %rd3;
	ld.param.u32 	%r19, [__cudaparmf6__Z33ConvertPixelTwo_422_8u_To_444_32f6uchar4S_14IR_PixelFormatP6float4S2_S0_];
	mov.s32 	%r20, %r19;
	.loc	22	60	0
	cvt.u8.u32 	%r21, %r2;
	cvt.rn.f32.u32 	%f1, %r21;
	mov.f32 	%f2, %f1;
	cvt.u8.u32 	%r22, %r4;
	cvt.rn.f32.u32 	%f3, %r22;
	mov.f32 	%f4, %f3;
	cvt.u8.u32 	%r23, %r6;
	cvt.rn.f32.u32 	%f5, %r23;
	mov.f32 	%f6, %f5;
	and.b32 	%r24, %r18, 4096;
	cvt.u8.u32 	%r25, %r8;
	mov.s32 	%r26, 0;
	setp.ne.s32 	%p1, %r24, %r26;
	cvt.rn.f32.u32 	%f7, %r25;
	@!%p1 bra 	$Lt_38_687106;
	.loc	23	86	0
	mov.f32 	%f2, %f3;
	.loc	23	87	0
	mov.f32 	%f4, %f1;
	.loc	22	38	0
	mov.f32 	%f6, %f7;
$Lt_38_687106:
	.loc	20	57	0
	and.b32 	%r27, %r18, 1536;
	and.b32 	%r28, %r18, 448;
	mov.s32 	%r29, 512;
	setp.eq.s32 	%p2, %r27, %r29;
	mov.s32 	%r30, 2565;
	mov.s32 	%r31, 2053;
	selp.s32 	%r32, %r30, %r31, %p2;
	or.b32 	%r33, %r32, %r28;
	and.b32 	%r34, %r33, 448;
	mov.s32 	%r35, 0;
	setp.eq.s32 	%p3, %r34, %r35;
	@!%p3 bra 	$Lt_38_432898;
	.loc	20	59	0
	mov.f32 	%f8, 0f437f0000;     	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_233;
$Lt_38_432898:
	.loc	20	61	0
	mov.u32 	%r36, 64;
	setp.ne.s32 	%p4, %r34, %r36;
	@%p4 bra 	$Lt_38_433154;
	.loc	20	63	0
	mov.f32 	%f8, 0f447fc000;     	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_233;
$Lt_38_433154:
	.loc	20	65	0
	mov.u32 	%r37, 128;
	setp.ne.s32 	%p5, %r34, %r37;
	@%p5 bra 	$Lt_38_433410;
	.loc	20	68	0
	mov.f32 	%f8, 0f47000000;     	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_233;
$Lt_38_433410:
	.loc	20	70	0
	mov.u32 	%r38, 192;
	setp.ne.s32 	%p6, %r34, %r38;
	@%p6 bra 	$Lt_38_433666;
	.loc	20	72	0
	mov.f32 	%f8, 0fbf800000;     	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_233;
$Lt_38_433666:
	.loc	20	76	0
	mov.f32 	%f8, 0f3f800000;     	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_233:
	.loc	20	469	0
	mov.f32 	%f9, %f6;
	mov.f32 	%f10, %f9;
	mov.f32 	%f11, %f2;
	mov.f32 	%f12, %f11;
	mov.f32 	%f13, %f4;
	mov.f32 	%f14, %f8;
	and.b32 	%r39, %r33, 4096;
	mov.s32 	%r40, 0;
	setp.eq.s32 	%p7, %r39, %r40;
	@!%p7 bra 	$Lt_38_687618;
	.loc	20	473	0
	mov.f32 	%f10, %f14;
	mov.f32 	%f12, %f13;
	mov.f32 	%f13, %f11;
	mov.f32 	%f14, %f9;
$Lt_38_687618:
	.loc	20	476	0
	and.b32 	%r41, %r20, 448;
	mov.s32 	%r42, %r20;
	mov.s32 	%r43, 256;
	setp.ne.s32 	%p8, %r41, %r43;
	and.b32 	%r44, %r42, 1;
	selp.s32 	%r45, 1, 0, %p8;
	and.b32 	%r46, %r44, %r45;
	mov.s32 	%r47, %r33;
	mov.s32 	%r48, 256;
	setp.ne.s32 	%p9, %r34, %r48;
	and.b32 	%r49, %r47, 1;
	selp.s32 	%r50, 1, 0, %p9;
	and.b32 	%r51, %r49, %r50;
	mov.u32 	%r52, 0;
	setp.eq.s32 	%p10, %r51, %r52;
	@%p10 bra 	$Lt_38_769538;
	mov.u32 	%r53, 0;
	setp.ne.s32 	%p11, %r46, %r53;
	@%p11 bra 	$Lt_38_769538;
	.loc	20	57	0
	@!%p3 bra 	$Lt_38_434178;
	.loc	20	59	0
	mov.f32 	%f15, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_231;
$Lt_38_434178:
	.loc	20	61	0
	mov.u32 	%r54, 64;
	setp.ne.s32 	%p12, %r34, %r54;
	@%p12 bra 	$Lt_38_434434;
	.loc	20	63	0
	mov.f32 	%f15, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_231;
$Lt_38_434434:
	.loc	20	65	0
	mov.u32 	%r55, 128;
	setp.ne.s32 	%p13, %r34, %r55;
	@%p13 bra 	$Lt_38_434690;
	.loc	20	68	0
	mov.f32 	%f15, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_231;
$Lt_38_434690:
	.loc	20	70	0
	mov.u32 	%r56, 192;
	setp.ne.s32 	%p14, %r34, %r56;
	@%p14 bra 	$Lt_38_434946;
	.loc	20	72	0
	mov.f32 	%f15, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_231;
$Lt_38_434946:
	.loc	20	76	0
	mov.f32 	%f15, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_231:
	.loc	20	118	0
	and.b32 	%r57, %r33, 2048;
	mov.s32 	%r58, 0;
	setp.ne.s32 	%p15, %r57, %r58;
	@!%p15 bra 	$Lt_38_688386;
	.loc	20	100	0
	ld.const.f32 	%f16, [kYCbCrOffset+0];
	bra.uni 	$Lt_38_688130;
$Lt_38_688386:
	ld.const.f32 	%f16, [kYCbCrFullRangeOffset+0];
$Lt_38_688130:
	.loc	20	118	0
	@!%p15 bra 	$Lt_38_688898;
	.loc	20	100	0
	ld.const.f32 	%f17, [kYCbCrOffset+4];
	bra.uni 	$Lt_38_688642;
$Lt_38_688898:
	ld.const.f32 	%f17, [kYCbCrFullRangeOffset+4];
$Lt_38_688642:
	.loc	20	118	0
	@!%p15 bra 	$Lt_38_689410;
	.loc	20	100	0
	ld.const.f32 	%f18, [kYCbCrOffset+8];
	bra.uni 	$Lt_38_689154;
$Lt_38_689410:
	ld.const.f32 	%f18, [kYCbCrFullRangeOffset+8];
$Lt_38_689154:
	.loc	20	478	0
	mov.f32 	%f19, 0f437f0000;    	// 255
	div.approx.ftz.f32 	%f20, %f15, %f19;
	mul.ftz.f32 	%f21, %f20, %f16;
	sub.ftz.f32 	%f12, %f12, %f21;
	mul.ftz.f32 	%f22, %f20, %f17;
	sub.ftz.f32 	%f13, %f13, %f22;
	mul.ftz.f32 	%f23, %f20, %f18;
	sub.ftz.f32 	%f14, %f14, %f23;
$Lt_38_769538:
$Lt_38_69634:
	.loc	20	481	0
	and.b32 	%r59, %r20, 2;
	and.b32 	%r60, %r33, 2;
	mov.s32 	%r61, 0;
	setp.ne.s32 	%p16, %r60, %r61;
	@!%p16 bra 	$Lt_38_770050;
	mov.u32 	%r62, 0;
	setp.ne.s32 	%p17, %r59, %r62;
	@%p17 bra 	$Lt_38_770050;
	.loc	20	483	0
	mov.f32 	%f24, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p18, %f12, %f24;
	@!%p18 bra 	$Lt_38_689666;
	.loc	20	372	0
	neg.ftz.f32 	%f25, %f12;
	lg2.approx.ftz.f32 	%f26, %f25;
	mov.f32 	%f27, 0f3ee66666;    	// 0.45
	mul.ftz.f32 	%f28, %f26, %f27;
	ex2.approx.ftz.f32 	%f29, %f28;
	neg.ftz.f32 	%f30, %f29;
	bra.uni 	$LDWendi___log2f_216_227;
$Lt_38_689666:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f31, %f12;
	mov.f32 	%f32, 0f3ee66666;    	// 0.45
	mul.ftz.f32 	%f33, %f31, %f32;
	ex2.approx.ftz.f32 	%f30, %f33;
$LDWendi___log2f_216_227:
	.loc	20	483	0
	mov.f32 	%f34, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p19, %f13, %f34;
	@!%p19 bra 	$Lt_38_690178;
	.loc	20	372	0
	neg.ftz.f32 	%f35, %f13;
	lg2.approx.ftz.f32 	%f36, %f35;
	mov.f32 	%f37, 0f3ee66666;    	// 0.45
	mul.ftz.f32 	%f38, %f36, %f37;
	ex2.approx.ftz.f32 	%f39, %f38;
	neg.ftz.f32 	%f40, %f39;
	bra.uni 	$LDWendi___log2f_216_225;
$Lt_38_690178:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f41, %f13;
	mov.f32 	%f42, 0f3ee66666;    	// 0.45
	mul.ftz.f32 	%f43, %f41, %f42;
	ex2.approx.ftz.f32 	%f40, %f43;
$LDWendi___log2f_216_225:
	.loc	20	483	0
	mov.f32 	%f44, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p20, %f14, %f44;
	@!%p20 bra 	$Lt_38_690690;
	.loc	20	372	0
	neg.ftz.f32 	%f45, %f14;
	lg2.approx.ftz.f32 	%f46, %f45;
	mov.f32 	%f47, 0f3ee66666;    	// 0.45
	mul.ftz.f32 	%f48, %f46, %f47;
	ex2.approx.ftz.f32 	%f49, %f48;
	neg.ftz.f32 	%f50, %f49;
	bra.uni 	$LDWendi___log2f_216_223;
$Lt_38_690690:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f51, %f14;
	mov.f32 	%f52, 0f3ee66666;    	// 0.45
	mul.ftz.f32 	%f53, %f51, %f52;
	ex2.approx.ftz.f32 	%f50, %f53;
$LDWendi___log2f_216_223:
	.loc	20	483	0
	mov.f32 	%f12, %f30;
	mov.f32 	%f13, %f40;
	mov.f32 	%f14, %f50;
$Lt_38_770050:
$Lt_38_73474:
	.loc	20	486	0
	and.b32 	%r63, %r20, 1;
	and.b32 	%r64, %r20, 1536;
	and.b32 	%r65, %r33, 1;
	and.b32 	%r66, %r33, 1536;
	setp.ne.s32 	%p21, %r63, %r65;
	setp.ne.s32 	%p22, %r64, %r66;
	selp.s32 	%r67, 1, 0, %p21;
	selp.s32 	%r68, 1, 0, %p22;
	or.b32 	%r69, %r67, %r68;
	mov.u32 	%r70, 0;
	setp.ne.s32 	%p23, %r69, %r70;
	@%p23 bra 	$Lt_38_124162;
	setp.eq.s32 	%p24, %r46, %r51;
	@%p24 bra 	$Lt_38_124418;
$Lt_38_124162:
	.loc	20	490	0
	mov.u32 	%r71, 0;
	setp.ne.s32 	%p25, %r65, %r71;
	@%p25 bra 	$Lt_38_691458;
	mov.s32 	%r72, 256;
	setp.eq.s32 	%p26, %r41, %r72;
	mov.u32 	%r73, 256;
	setp.ne.s32 	%p27, %r34, %r73;
	@%p27 bra 	$Lt_38_691970;
	.loc	20	137	0
	mov.s32 	%r74, 512;
	setp.eq.s32 	%p28, %r64, %r74;
	@!%p26 bra 	$Lt_38_438274;
	.loc	20	139	0
	@!%p28 bra 	$Lt_38_438530;
	.loc	20	141	0
	cvta.global.u64 	%rd5, kRGB32f_To_709YPbPr;
	mov.s64 	%rd6, %rd5;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_221;
$Lt_38_438530:
	.loc	20	145	0
	cvta.global.u64 	%rd7, kRGB32f_To_601YPbPr;
	mov.s64 	%rd6, %rd7;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_221;
$Lt_38_438274:
	.loc	20	150	0
	@!%p28 bra 	$Lt_38_438786;
	.loc	20	152	0
	cvta.global.u64 	%rd8, kRGB32f_To_709YCbCr;
	mov.s64 	%rd6, %rd8;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_221;
$Lt_38_438786:
	.loc	20	154	0
	and.b32 	%r75, %r20, 2048;
	mov.u32 	%r76, 0;
	setp.ne.s32 	%p29, %r75, %r76;
	@%p29 bra 	$Lt_38_439042;
	.loc	20	156	0
	cvta.global.u64 	%rd9, kRGB32f_To_601YCbCrFullRange;
	mov.s64 	%rd6, %rd9;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_221;
$Lt_38_439042:
	.loc	20	160	0
	cvta.global.u64 	%rd10, kRGB32f_To_601YCbCr;
	mov.s64 	%rd6, %rd10;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_221;
$Lt_38_691970:
	@!%p26 bra 	$Lt_38_692482;
	bra.uni 	$Lt_38_691202;
$Lt_38_692482:
	.loc	20	179	0
	mov.u32 	%r77, 512;
	setp.ne.s32 	%p30, %r64, %r77;
	@%p30 bra 	$Lt_38_439810;
	.loc	20	181	0
	cvta.global.u64 	%rd11, kRGB8u_To_709YCbCr;
	mov.s64 	%rd6, %rd11;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_221;
$Lt_38_439810:
	.loc	20	183	0
	and.b32 	%r78, %r20, 2048;
	mov.u32 	%r79, 0;
	setp.ne.s32 	%p31, %r78, %r79;
	@%p31 bra 	$Lt_38_440066;
	.loc	20	185	0
	cvta.global.u64 	%rd12, kRGB8u_To_601YCbCrFullRange;
	mov.s64 	%rd6, %rd12;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_221;
$Lt_38_440066:
	.loc	20	189	0
	cvta.global.u64 	%rd13, kRGB8u_To_601YCbCr;
	mov.s64 	%rd6, %rd13;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_221;
$Lt_38_691458:
	mov.s32 	%r80, 0;
	setp.eq.s32 	%p32, %r63, %r80;
	mov.u32 	%r81, 512;
	setp.ne.s32 	%p33, %r66, %r81;
	@%p33 bra 	$Lt_38_692994;
	mov.s32 	%r82, 256;
	setp.eq.s32 	%p34, %r34, %r82;
	@!%p32 bra 	$Lt_38_693506;
	mov.s32 	%r83, 256;
	setp.eq.s32 	%p26, %r41, %r83;
	@!%p34 bra 	$Lt_38_694018;
	@!%p26 bra 	$Lt_38_691202;
	.loc	20	202	0
	cvta.global.u64 	%rd14, k709YPbPr_To_RGB32f;
	mov.s64 	%rd6, %rd14;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_221;
$Lt_38_694018:
	.loc	20	211	0
	@!%p26 bra 	$Lt_38_441346;
	.loc	20	213	0
	cvta.global.u64 	%rd15, k709YCbCr_To_RGB32f;
	mov.s64 	%rd6, %rd15;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_221;
$Lt_38_441346:
	.loc	20	217	0
	cvta.global.u64 	%rd16, k709YCbCr_To_RGB8u;
	mov.s64 	%rd6, %rd16;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_221;
$Lt_38_693506:
	@!%p34 bra 	$Lt_38_695042;
	bra.uni 	$Lt_38_691202;
$Lt_38_695042:
	mov.s32 	%r84, 256;
	set.eq.u32.s32 	%r85, %r41, %r84;
	neg.s32 	%r86, %r85;
	and.b32 	%r87, %r20, 2048;
	mov.s32 	%r88, 0;
	set.eq.u32.s32 	%r89, %r87, %r88;
	neg.s32 	%r90, %r89;
	or.b32 	%r91, %r86, %r90;
	mov.u32 	%r92, 0;
	setp.eq.s32 	%p35, %r91, %r92;
	@%p35 bra 	$Lt_38_695554;
	bra.uni 	$Lt_38_691202;
$Lt_38_695554:
	cvta.global.u64 	%rd17, k709YCbCr_To_601YCbCr;
	.loc	20	250	0
	mov.s64 	%rd6, %rd17;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_221;
$Lt_38_692994:
	and.b32 	%r57, %r33, 2048;
	mov.s32 	%r93, 0;
	setp.eq.s32 	%p36, %r57, %r93;
	mov.pred 	%p37, %p36;
	mov.pred 	%p38, %p39;
	@!%p36 bra 	$Lt_38_696066;
	@!%p32 bra 	$Lt_38_691202;
	.loc	20	259	0
	mov.u32 	%r94, 256;
	setp.ne.s32 	%p40, %r41, %r94;
	@%p40 bra 	$Lt_38_442882;
	.loc	20	261	0
	cvta.global.u64 	%rd18, k601YCbCrFullRange_To_RGB32f;
	mov.s64 	%rd6, %rd18;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_221;
$Lt_38_442882:
	.loc	20	265	0
	cvta.global.u64 	%rd19, k601YCbCrFullRange_To_RGB8u;
	mov.s64 	%rd6, %rd19;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_221;
$Lt_38_696066:
	mov.s32 	%r95, 256;
	setp.eq.s32 	%p34, %r34, %r95;
	@!%p32 bra 	$Lt_38_697090;
	mov.s32 	%r96, 256;
	setp.eq.s32 	%p26, %r41, %r96;
	@!%p34 bra 	$Lt_38_697602;
	@!%p26 bra 	$Lt_38_691202;
	.loc	20	302	0
	cvta.global.u64 	%rd20, k601YPbPr_To_RGB32f;
	mov.s64 	%rd6, %rd20;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_221;
$Lt_38_697602:
	.loc	20	311	0
	@!%p26 bra 	$Lt_38_444674;
	.loc	20	313	0
	cvta.global.u64 	%rd21, k601YCbCr_To_RGB32f;
	mov.s64 	%rd6, %rd21;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_221;
$Lt_38_444674:
	.loc	20	317	0
	cvta.global.u64 	%rd22, k601YCbCr_To_RGB8u;
	mov.s64 	%rd6, %rd22;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_221;
$Lt_38_697090:
	@!%p34 bra 	$Lt_38_698626;
	bra.uni 	$Lt_38_691202;
$Lt_38_698626:
	selp.s32 	%r97, 1, 0, %p37;
	mov.s32 	%r98, 256;
	set.eq.u32.s32 	%r99, %r41, %r98;
	neg.s32 	%r100, %r99;
	or.b32 	%r101, %r97, %r100;
	mov.u32 	%r102, 0;
	setp.eq.s32 	%p41, %r101, %r102;
	@%p41 bra 	$Lt_38_699138;
	bra.uni 	$Lt_38_691202;
$Lt_38_699138:
	cvta.global.u64 	%rd23, k601YCbCr_To_709YCbCr;
	.loc	20	350	0
	mov.s64 	%rd6, %rd23;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_221;
$Lt_38_691202:
	.loc	20	355	0
	mov.u64 	%rd6, 0;
$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_221:
	.loc	20	490	0
	cvta.global.u64 	%rd7, kRGB32f_To_601YPbPr;
	cvta.global.u64 	%rd10, kRGB32f_To_601YCbCr;
	cvta.global.u64 	%rd13, kRGB8u_To_601YCbCr;
	cvta.global.u64 	%rd12, kRGB8u_To_601YCbCrFullRange;
	cvta.global.u64 	%rd9, kRGB32f_To_601YCbCrFullRange;
	cvta.global.u64 	%rd5, kRGB32f_To_709YPbPr;
	cvta.global.u64 	%rd8, kRGB32f_To_709YCbCr;
	cvta.global.u64 	%rd11, kRGB8u_To_709YCbCr;
	cvta.global.u64 	%rd17, k709YCbCr_To_601YCbCr;
	cvta.global.u64 	%rd20, k601YPbPr_To_RGB32f;
	cvta.global.u64 	%rd21, k601YCbCr_To_RGB32f;
	cvta.global.u64 	%rd22, k601YCbCr_To_RGB8u;
	cvta.global.u64 	%rd19, k601YCbCrFullRange_To_RGB8u;
	cvta.global.u64 	%rd18, k601YCbCrFullRange_To_RGB32f;
	cvta.global.u64 	%rd14, k709YPbPr_To_RGB32f;
	cvta.global.u64 	%rd15, k709YCbCr_To_RGB32f;
	cvta.global.u64 	%rd16, k709YCbCr_To_RGB8u;
	cvta.global.u64 	%rd23, k601YCbCr_To_709YCbCr;
	ld.f32 	%f54, [%rd6+16];
	mul.ftz.f32 	%f55, %f54, %f13;
	ld.f32 	%f56, [%rd6+12];
	fma.rn.ftz.f32 	%f57, %f56, %f12, %f55;
	ld.f32 	%f58, [%rd6+20];
	fma.rn.ftz.f32 	%f59, %f58, %f14, %f57;
	ld.f32 	%f60, [%rd6+28];
	mul.ftz.f32 	%f61, %f60, %f13;
	ld.f32 	%f62, [%rd6+24];
	fma.rn.ftz.f32 	%f63, %f62, %f12, %f61;
	ld.f32 	%f64, [%rd6+32];
	fma.rn.ftz.f32 	%f65, %f64, %f14, %f63;
	ld.f32 	%f66, [%rd6+4];
	mul.ftz.f32 	%f67, %f66, %f13;
	ld.f32 	%f68, [%rd6+0];
	fma.rn.ftz.f32 	%f69, %f68, %f12, %f67;
	ld.f32 	%f70, [%rd6+8];
	fma.rn.ftz.f32 	%f12, %f70, %f14, %f69;
	mov.f32 	%f13, %f59;
	mov.f32 	%f14, %f65;
	setp.ne.s32 	%p42, %r41, %r34;
	@!%p42 bra 	$Lt_38_126722;
	.loc	20	494	0
	mov.s32 	%r103, 256;
	setp.eq.s32 	%p34, %r34, %r103;
	@!%p34 bra 	$L_38_677890;
	mov.s32 	%r104, 0;
	setp.eq.s32 	%p43, %r41, %r104;
	@%p43 bra 	$L_38_677634;
$L_38_677890:
	@!%p3 bra 	$Lt_38_771330;
	mov.u32 	%r105, 256;
	setp.ne.s32 	%p44, %r41, %r105;
	@%p44 bra 	$Lt_38_771330;
	mov.s32 	%r106, 0;
	setp.eq.s32 	%p43, %r41, %r106;
$L_38_677634:
	.loc	20	57	0
	@!%p43 bra 	$Lt_38_446210;
	.loc	20	59	0
	mov.f32 	%f71, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_219;
$Lt_38_446210:
	.loc	20	61	0
	mov.u32 	%r107, 64;
	setp.ne.s32 	%p45, %r41, %r107;
	@%p45 bra 	$Lt_38_446466;
	.loc	20	63	0
	mov.f32 	%f71, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_219;
$Lt_38_446466:
	.loc	20	65	0
	mov.u32 	%r108, 128;
	setp.ne.s32 	%p46, %r41, %r108;
	@%p46 bra 	$Lt_38_446722;
	.loc	20	68	0
	mov.f32 	%f71, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_219;
$Lt_38_446722:
	.loc	20	70	0
	mov.u32 	%r109, 192;
	setp.ne.s32 	%p47, %r41, %r109;
	@%p47 bra 	$Lt_38_446978;
	.loc	20	72	0
	mov.f32 	%f71, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_219;
$Lt_38_446978:
	.loc	20	76	0
	mov.f32 	%f71, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_219:
	.loc	20	57	0
	@!%p3 bra 	$Lt_38_447234;
	.loc	20	59	0
	mov.f32 	%f72, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_217;
$Lt_38_447234:
	.loc	20	61	0
	mov.u32 	%r110, 64;
	setp.ne.s32 	%p48, %r34, %r110;
	@%p48 bra 	$Lt_38_447490;
	.loc	20	63	0
	mov.f32 	%f72, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_217;
$Lt_38_447490:
	.loc	20	65	0
	mov.u32 	%r111, 128;
	setp.ne.s32 	%p49, %r34, %r111;
	@%p49 bra 	$Lt_38_447746;
	.loc	20	68	0
	mov.f32 	%f72, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_217;
$Lt_38_447746:
	.loc	20	70	0
	mov.u32 	%r112, 192;
	setp.ne.s32 	%p50, %r34, %r112;
	@%p50 bra 	$Lt_38_448002;
	.loc	20	72	0
	mov.f32 	%f72, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_217;
$Lt_38_448002:
	.loc	20	76	0
	mov.f32 	%f72, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_217:
	.loc	20	498	0
	div.approx.ftz.f32 	%f73, %f71, %f72;
	mul.ftz.f32 	%f10, %f10, %f73;
	bra.uni 	$Lt_38_126722;
$Lt_38_771330:
$L_38_677378:
	.loc	20	500	0
	@!%p34 bra 	$L_38_678914;
	@%p8 bra 	$L_38_678658;
$L_38_678914:
	@!%p9 bra 	$Lt_38_772354;
	mov.u32 	%r113, 256;
	setp.ne.s32 	%p51, %r41, %r113;
	@%p51 bra 	$Lt_38_772354;
$L_38_678658:
	.loc	20	57	0
	mov.u32 	%r114, 0;
	setp.ne.s32 	%p52, %r41, %r114;
	@%p52 bra 	$Lt_38_448514;
	.loc	20	59	0
	mov.f32 	%f71, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_215;
$Lt_38_448514:
	.loc	20	61	0
	mov.u32 	%r115, 64;
	setp.ne.s32 	%p53, %r41, %r115;
	@%p53 bra 	$Lt_38_448770;
	.loc	20	63	0
	mov.f32 	%f71, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_215;
$Lt_38_448770:
	.loc	20	65	0
	mov.u32 	%r116, 128;
	setp.ne.s32 	%p54, %r41, %r116;
	@%p54 bra 	$Lt_38_449026;
	.loc	20	68	0
	mov.f32 	%f71, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_215;
$Lt_38_449026:
	.loc	20	70	0
	mov.u32 	%r117, 192;
	setp.ne.s32 	%p55, %r41, %r117;
	@%p55 bra 	$Lt_38_449282;
	.loc	20	72	0
	mov.f32 	%f71, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_215;
$Lt_38_449282:
	.loc	20	76	0
	mov.f32 	%f71, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_215:
	.loc	20	57	0
	@!%p3 bra 	$Lt_38_449538;
	.loc	20	59	0
	mov.f32 	%f72, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_213;
$Lt_38_449538:
	.loc	20	61	0
	mov.u32 	%r118, 64;
	setp.ne.s32 	%p56, %r34, %r118;
	@%p56 bra 	$Lt_38_449794;
	.loc	20	63	0
	mov.f32 	%f72, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_213;
$Lt_38_449794:
	.loc	20	65	0
	mov.u32 	%r119, 128;
	setp.ne.s32 	%p57, %r34, %r119;
	@%p57 bra 	$Lt_38_450050;
	.loc	20	68	0
	mov.f32 	%f72, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_213;
$Lt_38_450050:
	.loc	20	70	0
	mov.u32 	%r120, 192;
	setp.ne.s32 	%p58, %r34, %r120;
	@%p58 bra 	$Lt_38_450306;
	.loc	20	72	0
	mov.f32 	%f72, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_213;
$Lt_38_450306:
	.loc	20	76	0
	mov.f32 	%f72, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_213:
	.loc	20	504	0
	div.approx.ftz.f32 	%f74, %f71, %f72;
	mul.ftz.f32 	%f10, %f10, %f74;
	.loc	20	57	0
	@!%p3 bra 	$Lt_38_451586;
	.loc	20	59	0
	mov.f32 	%f72, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_209;
$Lt_38_451586:
	.loc	20	61	0
	mov.u32 	%r121, 64;
	setp.ne.s32 	%p59, %r34, %r121;
	@%p59 bra 	$Lt_38_451842;
	.loc	20	63	0
	mov.f32 	%f72, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_209;
$Lt_38_451842:
	.loc	20	65	0
	mov.u32 	%r122, 128;
	setp.ne.s32 	%p60, %r34, %r122;
	@%p60 bra 	$Lt_38_452098;
	.loc	20	68	0
	mov.f32 	%f72, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_209;
$Lt_38_452098:
	.loc	20	70	0
	mov.u32 	%r123, 192;
	setp.ne.s32 	%p61, %r34, %r123;
	@%p61 bra 	$Lt_38_452354;
	.loc	20	72	0
	mov.f32 	%f72, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_209;
$Lt_38_452354:
	.loc	20	76	0
	mov.f32 	%f72, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_209:
	.loc	20	505	0
	mov.f32 	%f75, 0f437f0000;    	// 255
	div.approx.ftz.f32 	%f76, %f75, %f72;
	mul.ftz.f32 	%f12, %f12, %f76;
	.loc	20	57	0
	@!%p3 bra 	$Lt_38_453634;
	.loc	20	59	0
	mov.f32 	%f72, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_205;
$Lt_38_453634:
	.loc	20	61	0
	mov.u32 	%r124, 64;
	setp.ne.s32 	%p62, %r34, %r124;
	@%p62 bra 	$Lt_38_453890;
	.loc	20	63	0
	mov.f32 	%f72, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_205;
$Lt_38_453890:
	.loc	20	65	0
	mov.u32 	%r125, 128;
	setp.ne.s32 	%p63, %r34, %r125;
	@%p63 bra 	$Lt_38_454146;
	.loc	20	68	0
	mov.f32 	%f72, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_205;
$Lt_38_454146:
	.loc	20	70	0
	mov.u32 	%r126, 192;
	setp.ne.s32 	%p64, %r34, %r126;
	@%p64 bra 	$Lt_38_454402;
	.loc	20	72	0
	mov.f32 	%f72, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_205;
$Lt_38_454402:
	.loc	20	76	0
	mov.f32 	%f72, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_205:
	.loc	20	506	0
	mov.f32 	%f77, 0f437f0000;    	// 255
	div.approx.ftz.f32 	%f78, %f77, %f72;
	mul.ftz.f32 	%f13, %f59, %f78;
	.loc	20	57	0
	@!%p3 bra 	$Lt_38_455682;
	.loc	20	59	0
	mov.f32 	%f72, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_201;
$Lt_38_455682:
	.loc	20	61	0
	mov.u32 	%r127, 64;
	setp.ne.s32 	%p65, %r34, %r127;
	@%p65 bra 	$Lt_38_455938;
	.loc	20	63	0
	mov.f32 	%f72, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_201;
$Lt_38_455938:
	.loc	20	65	0
	mov.u32 	%r128, 128;
	setp.ne.s32 	%p66, %r34, %r128;
	@%p66 bra 	$Lt_38_456194;
	.loc	20	68	0
	mov.f32 	%f72, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_201;
$Lt_38_456194:
	.loc	20	70	0
	mov.u32 	%r129, 192;
	setp.ne.s32 	%p67, %r34, %r129;
	@%p67 bra 	$Lt_38_456450;
	.loc	20	72	0
	mov.f32 	%f72, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_201;
$Lt_38_456450:
	.loc	20	76	0
	mov.f32 	%f72, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_201:
	.loc	20	507	0
	mov.f32 	%f79, 0f437f0000;    	// 255
	div.approx.ftz.f32 	%f80, %f79, %f72;
	mul.ftz.f32 	%f14, %f65, %f80;
	bra.uni 	$Lt_38_126722;
$Lt_38_772354:
$L_38_678402:
	.loc	20	57	0
	@!%p3 bra 	$Lt_38_456706;
	.loc	20	59	0
	mov.f32 	%f71, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_199;
$Lt_38_456706:
	.loc	20	61	0
	mov.u32 	%r130, 64;
	setp.ne.s32 	%p68, %r34, %r130;
	@%p68 bra 	$Lt_38_456962;
	.loc	20	63	0
	mov.f32 	%f71, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_199;
$Lt_38_456962:
	.loc	20	65	0
	mov.u32 	%r131, 128;
	setp.ne.s32 	%p69, %r34, %r131;
	@%p69 bra 	$Lt_38_457218;
	.loc	20	68	0
	mov.f32 	%f71, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_199;
$Lt_38_457218:
	.loc	20	70	0
	mov.u32 	%r132, 192;
	setp.ne.s32 	%p70, %r34, %r132;
	@%p70 bra 	$Lt_38_457474;
	.loc	20	72	0
	mov.f32 	%f71, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_199;
$Lt_38_457474:
	.loc	20	76	0
	mov.f32 	%f71, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_199:
	.loc	20	511	0
	mov.f32 	%f81, 0f437f0000;    	// 255
	div.approx.ftz.f32 	%f82, %f71, %f81;
	mul.ftz.f32 	%f10, %f82, %f10;
	mul.ftz.f32 	%f12, %f82, %f12;
	mul.ftz.f32 	%f13, %f82, %f59;
	mul.ftz.f32 	%f14, %f82, %f65;
	bra.uni 	$Lt_38_126722;
$Lt_38_124418:
	.loc	20	486	0
	setp.ne.s32 	%p42, %r41, %r34;
	@!%p42 bra 	$Lt_38_699906;
	.loc	20	57	0
	mov.u32 	%r133, 0;
	setp.ne.s32 	%p71, %r41, %r133;
	@%p71 bra 	$Lt_38_459010;
	.loc	20	59	0
	mov.f32 	%f71, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_195;
$Lt_38_459010:
	.loc	20	61	0
	mov.u32 	%r134, 64;
	setp.ne.s32 	%p72, %r41, %r134;
	@%p72 bra 	$Lt_38_459266;
	.loc	20	63	0
	mov.f32 	%f71, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_195;
$Lt_38_459266:
	.loc	20	65	0
	mov.u32 	%r135, 128;
	setp.ne.s32 	%p73, %r41, %r135;
	@%p73 bra 	$Lt_38_459522;
	.loc	20	68	0
	mov.f32 	%f71, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_195;
$Lt_38_459522:
	.loc	20	70	0
	mov.u32 	%r136, 192;
	setp.ne.s32 	%p74, %r41, %r136;
	@%p74 bra 	$Lt_38_459778;
	.loc	20	72	0
	mov.f32 	%f71, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_195;
$Lt_38_459778:
	.loc	20	76	0
	mov.f32 	%f71, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_195:
	.loc	20	57	0
	@!%p3 bra 	$Lt_38_460034;
	.loc	20	59	0
	mov.f32 	%f72, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_193;
$Lt_38_460034:
	.loc	20	61	0
	mov.u32 	%r137, 64;
	setp.ne.s32 	%p75, %r34, %r137;
	@%p75 bra 	$Lt_38_460290;
	.loc	20	63	0
	mov.f32 	%f72, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_193;
$Lt_38_460290:
	.loc	20	65	0
	mov.u32 	%r138, 128;
	setp.ne.s32 	%p76, %r34, %r138;
	@%p76 bra 	$Lt_38_460546;
	.loc	20	68	0
	mov.f32 	%f72, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_193;
$Lt_38_460546:
	.loc	20	70	0
	mov.u32 	%r139, 192;
	setp.ne.s32 	%p77, %r34, %r139;
	@%p77 bra 	$Lt_38_460802;
	.loc	20	72	0
	mov.f32 	%f72, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_193;
$Lt_38_460802:
	.loc	20	76	0
	mov.f32 	%f72, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_193:
	.loc	20	517	0
	div.approx.ftz.f32 	%f83, %f71, %f72;
	mul.ftz.f32 	%f10, %f83, %f10;
	mul.ftz.f32 	%f12, %f83, %f12;
	mul.ftz.f32 	%f13, %f83, %f13;
	mul.ftz.f32 	%f14, %f83, %f14;
$Lt_38_699906:
	cvta.global.u64 	%rd23, k601YCbCr_To_709YCbCr;
	cvta.global.u64 	%rd16, k709YCbCr_To_RGB8u;
	cvta.global.u64 	%rd15, k709YCbCr_To_RGB32f;
	cvta.global.u64 	%rd14, k709YPbPr_To_RGB32f;
	cvta.global.u64 	%rd18, k601YCbCrFullRange_To_RGB32f;
	cvta.global.u64 	%rd19, k601YCbCrFullRange_To_RGB8u;
	cvta.global.u64 	%rd22, k601YCbCr_To_RGB8u;
	cvta.global.u64 	%rd21, k601YCbCr_To_RGB32f;
	cvta.global.u64 	%rd20, k601YPbPr_To_RGB32f;
	cvta.global.u64 	%rd17, k709YCbCr_To_601YCbCr;
	cvta.global.u64 	%rd11, kRGB8u_To_709YCbCr;
	cvta.global.u64 	%rd8, kRGB32f_To_709YCbCr;
	cvta.global.u64 	%rd5, kRGB32f_To_709YPbPr;
	cvta.global.u64 	%rd9, kRGB32f_To_601YCbCrFullRange;
	cvta.global.u64 	%rd12, kRGB8u_To_601YCbCrFullRange;
	cvta.global.u64 	%rd13, kRGB8u_To_601YCbCr;
	cvta.global.u64 	%rd10, kRGB32f_To_601YCbCr;
	cvta.global.u64 	%rd7, kRGB32f_To_601YPbPr;
$Lt_38_126722:
	.loc	20	520	0
	mov.s32 	%r140, 0;
	setp.ne.s32 	%p78, %r59, %r140;
	@!%p78 bra 	$Lt_38_772866;
	mov.u32 	%r141, 0;
	setp.ne.s32 	%p79, %r60, %r141;
	@%p79 bra 	$Lt_38_772866;
	.loc	20	522	0
	mov.f32 	%f84, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p80, %f12, %f84;
	@!%p80 bra 	$Lt_38_700418;
	.loc	20	372	0
	neg.ftz.f32 	%f85, %f12;
	lg2.approx.ftz.f32 	%f86, %f85;
	mov.f32 	%f87, 0f400e38e4;    	// 2.22222
	mul.ftz.f32 	%f88, %f86, %f87;
	ex2.approx.ftz.f32 	%f89, %f88;
	neg.ftz.f32 	%f90, %f89;
	bra.uni 	$LDWendi___log2f_216_191;
$Lt_38_700418:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f91, %f12;
	mov.f32 	%f92, 0f400e38e4;    	// 2.22222
	mul.ftz.f32 	%f93, %f91, %f92;
	ex2.approx.ftz.f32 	%f90, %f93;
$LDWendi___log2f_216_191:
	.loc	20	522	0
	mov.f32 	%f94, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p81, %f13, %f94;
	@!%p81 bra 	$Lt_38_700930;
	.loc	20	372	0
	neg.ftz.f32 	%f95, %f13;
	lg2.approx.ftz.f32 	%f96, %f95;
	mov.f32 	%f97, 0f400e38e4;    	// 2.22222
	mul.ftz.f32 	%f98, %f96, %f97;
	ex2.approx.ftz.f32 	%f99, %f98;
	neg.ftz.f32 	%f100, %f99;
	bra.uni 	$LDWendi___log2f_216_189;
$Lt_38_700930:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f101, %f13;
	mov.f32 	%f102, 0f400e38e4;   	// 2.22222
	mul.ftz.f32 	%f103, %f101, %f102;
	ex2.approx.ftz.f32 	%f100, %f103;
$LDWendi___log2f_216_189:
	.loc	20	522	0
	mov.f32 	%f104, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p82, %f14, %f104;
	@!%p82 bra 	$Lt_38_701442;
	.loc	20	372	0
	neg.ftz.f32 	%f105, %f14;
	lg2.approx.ftz.f32 	%f106, %f105;
	mov.f32 	%f107, 0f400e38e4;   	// 2.22222
	mul.ftz.f32 	%f108, %f106, %f107;
	ex2.approx.ftz.f32 	%f109, %f108;
	neg.ftz.f32 	%f110, %f109;
	bra.uni 	$LDWendi___log2f_216_187;
$Lt_38_701442:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f111, %f14;
	mov.f32 	%f112, 0f400e38e4;   	// 2.22222
	mul.ftz.f32 	%f113, %f111, %f112;
	ex2.approx.ftz.f32 	%f110, %f113;
$LDWendi___log2f_216_187:
	.loc	20	522	0
	mov.f32 	%f12, %f90;
	mov.f32 	%f13, %f100;
	mov.f32 	%f14, %f110;
$Lt_38_772866:
$Lt_38_128770:
	.loc	20	525	0
	mov.u32 	%r142, 0;
	setp.eq.s32 	%p83, %r46, %r142;
	@%p83 bra 	$Lt_38_773378;
	mov.u32 	%r143, 0;
	setp.ne.s32 	%p84, %r51, %r143;
	@%p84 bra 	$Lt_38_773378;
	.loc	20	57	0
	mov.u32 	%r144, 0;
	setp.ne.s32 	%p85, %r41, %r144;
	@%p85 bra 	$Lt_38_461826;
	.loc	20	59	0
	mov.f32 	%f114, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_185;
$Lt_38_461826:
	.loc	20	61	0
	mov.u32 	%r145, 64;
	setp.ne.s32 	%p86, %r41, %r145;
	@%p86 bra 	$Lt_38_462082;
	.loc	20	63	0
	mov.f32 	%f114, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_185;
$Lt_38_462082:
	.loc	20	65	0
	mov.u32 	%r146, 128;
	setp.ne.s32 	%p87, %r41, %r146;
	@%p87 bra 	$Lt_38_462338;
	.loc	20	68	0
	mov.f32 	%f114, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_185;
$Lt_38_462338:
	.loc	20	70	0
	mov.u32 	%r147, 192;
	setp.ne.s32 	%p88, %r41, %r147;
	@%p88 bra 	$Lt_38_462594;
	.loc	20	72	0
	mov.f32 	%f114, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_185;
$Lt_38_462594:
	.loc	20	76	0
	mov.f32 	%f114, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_185:
	.loc	20	107	0
	and.b32 	%r148, %r20, 2048;
	mov.s32 	%r149, 0;
	setp.ne.s32 	%p89, %r148, %r149;
	@!%p89 bra 	$Lt_38_702210;
	.loc	20	100	0
	ld.const.f32 	%f115, [kYCbCrOffset+0];
	bra.uni 	$Lt_38_701954;
$Lt_38_702210:
	ld.const.f32 	%f115, [kYCbCrFullRangeOffset+0];
$Lt_38_701954:
	.loc	20	107	0
	@!%p89 bra 	$Lt_38_702722;
	.loc	20	100	0
	ld.const.f32 	%f116, [kYCbCrOffset+4];
	bra.uni 	$Lt_38_702466;
$Lt_38_702722:
	ld.const.f32 	%f116, [kYCbCrFullRangeOffset+4];
$Lt_38_702466:
	.loc	20	107	0
	@!%p89 bra 	$Lt_38_703234;
	.loc	20	100	0
	ld.const.f32 	%f117, [kYCbCrOffset+8];
	bra.uni 	$Lt_38_702978;
$Lt_38_703234:
	ld.const.f32 	%f117, [kYCbCrFullRangeOffset+8];
$Lt_38_702978:
	.loc	20	527	0
	mov.f32 	%f118, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f119, %f114, %f118;
	fma.rn.ftz.f32 	%f12, %f119, %f115, %f12;
	fma.rn.ftz.f32 	%f13, %f119, %f116, %f13;
	fma.rn.ftz.f32 	%f14, %f119, %f117, %f14;
$Lt_38_773378:
$Lt_38_135170:
	.loc	20	525	0
	and.b32 	%r150, %r20, 12;
	and.b32 	%r151, %r33, 12;
	setp.ne.s32 	%p90, %r150, %r151;
	@!%p90 bra 	$Lt_38_704002;
	.loc	20	532	0
	mov.u32 	%r152, 8;
	setp.ne.s32 	%p91, %r151, %r152;
	@%p91 bra 	$L_38_680194;
	mov.u32 	%r153, 12;
	setp.eq.s32 	%p92, %r150, %r153;
	@%p92 bra 	$Lt_38_774146;
$L_38_680194:
	mov.u32 	%r154, 12;
	setp.eq.s32 	%p93, %r151, %r154;
	@%p93 bra 	$Lt_38_774146;
	mov.u32 	%r155, 0;
	setp.ne.s32 	%p94, %r151, %r155;
	@%p94 bra 	$L_38_679426;
$Lt_38_774146:
$L_38_679682:
	.loc	20	57	0
	mov.u32 	%r156, 0;
	setp.ne.s32 	%p95, %r41, %r156;
	@%p95 bra 	$Lt_38_465154;
	.loc	20	59	0
	mov.f32 	%f120, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_181;
$Lt_38_465154:
	.loc	20	61	0
	mov.u32 	%r157, 64;
	setp.ne.s32 	%p96, %r41, %r157;
	@%p96 bra 	$Lt_38_465410;
	.loc	20	63	0
	mov.f32 	%f120, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_181;
$Lt_38_465410:
	.loc	20	65	0
	mov.u32 	%r158, 128;
	setp.ne.s32 	%p97, %r41, %r158;
	@%p97 bra 	$Lt_38_465666;
	.loc	20	68	0
	mov.f32 	%f120, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_181;
$Lt_38_465666:
	.loc	20	70	0
	mov.u32 	%r159, 192;
	setp.ne.s32 	%p98, %r41, %r159;
	@%p98 bra 	$Lt_38_465922;
	.loc	20	72	0
	mov.f32 	%f120, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_181;
$Lt_38_465922:
	.loc	20	76	0
	mov.f32 	%f120, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_181:
	.loc	20	536	0
	mov.f32 	%f10, %f120;
	bra.uni 	$Lt_38_704002;
$L_38_679426:
	.loc	20	540	0
	mov.s32 	%r160, 12;
	setp.eq.s32 	%p99, %r150, %r160;
	mov.s32 	%r161, 4;
	set.eq.u32.s32 	%r162, %r151, %r161;
	neg.s32 	%r163, %r162;
	selp.s32 	%r164, 1, 0, %p99;
	mov.s32 	%r165, 8;
	set.eq.u32.s32 	%r166, %r150, %r165;
	neg.s32 	%r167, %r166;
	or.b32 	%r168, %r164, %r167;
	and.b32 	%r169, %r163, %r168;
	mov.u32 	%r170, 0;
	setp.eq.s32 	%p100, %r169, %r170;
	@%p100 bra 	$Lt_38_704258;
	.loc	20	410	0
	mov.f32 	%f121, %f12;
	mov.f32 	%f122, %f121;
	mov.f32 	%f123, %f13;
	mov.f32 	%f124, %f123;
	mov.f32 	%f125, %f14;
	mov.f32 	%f126, %f125;
	.loc	20	57	0
	mov.s32 	%r171, 0;
	setp.eq.s32 	%p43, %r41, %r171;
	@!%p43 bra 	$Lt_38_466434;
	.loc	20	59	0
	mov.f32 	%f127, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_179;
$Lt_38_466434:
	.loc	20	61	0
	mov.u32 	%r172, 64;
	setp.ne.s32 	%p101, %r41, %r172;
	@%p101 bra 	$Lt_38_466690;
	.loc	20	63	0
	mov.f32 	%f127, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_179;
$Lt_38_466690:
	.loc	20	65	0
	mov.u32 	%r173, 128;
	setp.ne.s32 	%p102, %r41, %r173;
	@%p102 bra 	$Lt_38_466946;
	.loc	20	68	0
	mov.f32 	%f127, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_179;
$Lt_38_466946:
	.loc	20	70	0
	mov.u32 	%r174, 192;
	setp.ne.s32 	%p103, %r41, %r174;
	@%p103 bra 	$Lt_38_467202;
	.loc	20	72	0
	mov.f32 	%f127, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_179;
$Lt_38_467202:
	.loc	20	76	0
	mov.f32 	%f127, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_179:
	.loc	20	413	0
	mov.u32 	%r175, 0;
	setp.eq.s32 	%p104, %r46, %r175;
	@%p104 bra 	$Lt_38_704514;
	.loc	20	57	0
	@!%p43 bra 	$Lt_38_467714;
	.loc	20	59	0
	mov.f32 	%f128, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_177;
$Lt_38_467714:
	.loc	20	61	0
	mov.u32 	%r176, 64;
	setp.ne.s32 	%p105, %r41, %r176;
	@%p105 bra 	$Lt_38_467970;
	.loc	20	63	0
	mov.f32 	%f128, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_177;
$Lt_38_467970:
	.loc	20	65	0
	mov.u32 	%r177, 128;
	setp.ne.s32 	%p106, %r41, %r177;
	@%p106 bra 	$Lt_38_468226;
	.loc	20	68	0
	mov.f32 	%f128, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_177;
$Lt_38_468226:
	.loc	20	70	0
	mov.u32 	%r178, 192;
	setp.ne.s32 	%p107, %r41, %r178;
	@%p107 bra 	$Lt_38_468482;
	.loc	20	72	0
	mov.f32 	%f128, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_177;
$Lt_38_468482:
	.loc	20	76	0
	mov.f32 	%f128, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_177:
	.loc	20	118	0
	and.b32 	%r148, %r20, 2048;
	mov.s32 	%r179, 0;
	setp.ne.s32 	%p89, %r148, %r179;
	@!%p89 bra 	$Lt_38_705282;
	.loc	20	100	0
	ld.const.f32 	%f129, [kYCbCrOffset+0];
	bra.uni 	$Lt_38_705026;
$Lt_38_705282:
	ld.const.f32 	%f129, [kYCbCrFullRangeOffset+0];
$Lt_38_705026:
	.loc	20	118	0
	@!%p89 bra 	$Lt_38_705794;
	.loc	20	100	0
	ld.const.f32 	%f130, [kYCbCrOffset+4];
	bra.uni 	$Lt_38_705538;
$Lt_38_705794:
	ld.const.f32 	%f130, [kYCbCrFullRangeOffset+4];
$Lt_38_705538:
	.loc	20	118	0
	@!%p89 bra 	$Lt_38_706306;
	.loc	20	100	0
	ld.const.f32 	%f131, [kYCbCrOffset+8];
	bra.uni 	$Lt_38_706050;
$Lt_38_706306:
	ld.const.f32 	%f131, [kYCbCrFullRangeOffset+8];
$Lt_38_706050:
	.loc	20	415	0
	mov.f32 	%f132, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f133, %f128, %f132;
	mul.ftz.f32 	%f134, %f133, %f129;
	sub.ftz.f32 	%f122, %f121, %f134;
	mul.ftz.f32 	%f135, %f133, %f130;
	sub.ftz.f32 	%f124, %f123, %f135;
	mul.ftz.f32 	%f136, %f133, %f131;
	sub.ftz.f32 	%f126, %f125, %f136;
$Lt_38_704514:
	.loc	20	418	0
	rcp.approx.ftz.f32 	%f137, %f127;
	mul.ftz.f32 	%f138, %f137, %f10;
	mul.ftz.f32 	%f122, %f138, %f122;
	.loc	20	419	0
	mul.ftz.f32 	%f124, %f138, %f124;
	.loc	20	420	0
	mul.ftz.f32 	%f126, %f138, %f126;
	.loc	20	422	0
	mov.u32 	%r180, 0;
	setp.eq.s32 	%p108, %r46, %r180;
	@%p108 bra 	$Lt_38_706562;
	.loc	20	57	0
	@!%p43 bra 	$Lt_38_470786;
	.loc	20	59	0
	mov.f32 	%f139, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_173;
$Lt_38_470786:
	.loc	20	61	0
	mov.u32 	%r181, 64;
	setp.ne.s32 	%p109, %r41, %r181;
	@%p109 bra 	$Lt_38_471042;
	.loc	20	63	0
	mov.f32 	%f139, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_173;
$Lt_38_471042:
	.loc	20	65	0
	mov.u32 	%r182, 128;
	setp.ne.s32 	%p110, %r41, %r182;
	@%p110 bra 	$Lt_38_471298;
	.loc	20	68	0
	mov.f32 	%f139, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_173;
$Lt_38_471298:
	.loc	20	70	0
	mov.u32 	%r183, 192;
	setp.ne.s32 	%p111, %r41, %r183;
	@%p111 bra 	$Lt_38_471554;
	.loc	20	72	0
	mov.f32 	%f139, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_173;
$Lt_38_471554:
	.loc	20	76	0
	mov.f32 	%f139, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_173:
	.loc	20	107	0
	and.b32 	%r148, %r20, 2048;
	mov.s32 	%r184, 0;
	setp.ne.s32 	%p89, %r148, %r184;
	@!%p89 bra 	$Lt_38_707330;
	.loc	20	100	0
	ld.const.f32 	%f140, [kYCbCrOffset+0];
	bra.uni 	$Lt_38_707074;
$Lt_38_707330:
	ld.const.f32 	%f140, [kYCbCrFullRangeOffset+0];
$Lt_38_707074:
	.loc	20	107	0
	@!%p89 bra 	$Lt_38_707842;
	.loc	20	100	0
	ld.const.f32 	%f141, [kYCbCrOffset+4];
	bra.uni 	$Lt_38_707586;
$Lt_38_707842:
	ld.const.f32 	%f141, [kYCbCrFullRangeOffset+4];
$Lt_38_707586:
	.loc	20	107	0
	@!%p89 bra 	$Lt_38_708354;
	.loc	20	100	0
	ld.const.f32 	%f142, [kYCbCrOffset+8];
	bra.uni 	$Lt_38_708098;
$Lt_38_708354:
	ld.const.f32 	%f142, [kYCbCrFullRangeOffset+8];
$Lt_38_708098:
	.loc	20	424	0
	mov.f32 	%f143, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f144, %f139, %f143;
	fma.rn.ftz.f32 	%f122, %f144, %f140, %f122;
	fma.rn.ftz.f32 	%f124, %f144, %f141, %f124;
	fma.rn.ftz.f32 	%f126, %f144, %f142, %f126;
$Lt_38_706562:
	.loc	20	543	0
	mov.f32 	%f12, %f122;
	mov.f32 	%f13, %f124;
	mov.f32 	%f14, %f126;
	@!%p99 bra 	$Lt_38_704002;
	.loc	20	57	0
	@!%p43 bra 	$Lt_38_473858;
	.loc	20	59	0
	mov.f32 	%f120, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_169;
$Lt_38_473858:
	.loc	20	61	0
	mov.u32 	%r185, 64;
	setp.ne.s32 	%p112, %r41, %r185;
	@%p112 bra 	$Lt_38_474114;
	.loc	20	63	0
	mov.f32 	%f120, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_169;
$Lt_38_474114:
	.loc	20	65	0
	mov.u32 	%r186, 128;
	setp.ne.s32 	%p113, %r41, %r186;
	@%p113 bra 	$Lt_38_474370;
	.loc	20	68	0
	mov.f32 	%f120, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_169;
$Lt_38_474370:
	.loc	20	70	0
	mov.u32 	%r187, 192;
	setp.ne.s32 	%p114, %r41, %r187;
	@%p114 bra 	$Lt_38_474626;
	.loc	20	72	0
	mov.f32 	%f120, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_169;
$Lt_38_474626:
	.loc	20	76	0
	mov.f32 	%f120, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_169:
	.loc	20	546	0
	mov.f32 	%f10, %f120;
	bra.uni 	$Lt_38_704002;
$Lt_38_704258:
	.loc	20	433	0
	mov.f32 	%f145, %f10;
	mov.f32 	%f146, %f12;
	mov.f32 	%f147, %f146;
	mov.f32 	%f148, %f13;
	mov.f32 	%f149, %f148;
	mov.f32 	%f150, %f14;
	mov.f32 	%f151, %f150;
	.loc	20	435	0
	mov.u32 	%r188, 0;
	setp.eq.s32 	%p115, %r46, %r188;
	@%p115 bra 	$Lt_38_709122;
	.loc	20	57	0
	mov.u32 	%r189, 0;
	setp.ne.s32 	%p116, %r41, %r189;
	@%p116 bra 	$Lt_38_475138;
	.loc	20	59	0
	mov.f32 	%f152, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_167;
$Lt_38_475138:
	.loc	20	61	0
	mov.u32 	%r190, 64;
	setp.ne.s32 	%p117, %r41, %r190;
	@%p117 bra 	$Lt_38_475394;
	.loc	20	63	0
	mov.f32 	%f152, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_167;
$Lt_38_475394:
	.loc	20	65	0
	mov.u32 	%r191, 128;
	setp.ne.s32 	%p118, %r41, %r191;
	@%p118 bra 	$Lt_38_475650;
	.loc	20	68	0
	mov.f32 	%f152, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_167;
$Lt_38_475650:
	.loc	20	70	0
	mov.u32 	%r192, 192;
	setp.ne.s32 	%p119, %r41, %r192;
	@%p119 bra 	$Lt_38_475906;
	.loc	20	72	0
	mov.f32 	%f152, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_167;
$Lt_38_475906:
	.loc	20	76	0
	mov.f32 	%f152, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_167:
	.loc	20	118	0
	and.b32 	%r148, %r20, 2048;
	mov.s32 	%r193, 0;
	setp.ne.s32 	%p89, %r148, %r193;
	@!%p89 bra 	$Lt_38_709890;
	.loc	20	100	0
	ld.const.f32 	%f153, [kYCbCrOffset+0];
	bra.uni 	$Lt_38_709634;
$Lt_38_709890:
	ld.const.f32 	%f153, [kYCbCrFullRangeOffset+0];
$Lt_38_709634:
	.loc	20	118	0
	@!%p89 bra 	$Lt_38_710402;
	.loc	20	100	0
	ld.const.f32 	%f154, [kYCbCrOffset+4];
	bra.uni 	$Lt_38_710146;
$Lt_38_710402:
	ld.const.f32 	%f154, [kYCbCrFullRangeOffset+4];
$Lt_38_710146:
	.loc	20	118	0
	@!%p89 bra 	$Lt_38_710914;
	.loc	20	100	0
	ld.const.f32 	%f155, [kYCbCrOffset+8];
	bra.uni 	$Lt_38_710658;
$Lt_38_710914:
	ld.const.f32 	%f155, [kYCbCrFullRangeOffset+8];
$Lt_38_710658:
	.loc	20	437	0
	mov.f32 	%f156, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f157, %f152, %f156;
	mul.ftz.f32 	%f158, %f157, %f153;
	sub.ftz.f32 	%f147, %f146, %f158;
	mul.ftz.f32 	%f159, %f157, %f154;
	sub.ftz.f32 	%f149, %f148, %f159;
	mul.ftz.f32 	%f160, %f157, %f155;
	sub.ftz.f32 	%f151, %f150, %f160;
$Lt_38_709122:
	mov.f32 	%f161, 0fb70637bd;   	// -8e-006
	add.ftz.f32 	%f162, %f10, %f161;
	mov.f32 	%f163, 0f00000000;   	// 0
	setp.le.ftz.f32 	%p120, %f162, %f163;
	@!%p120 bra 	$Lt_38_711426;
	mov.f32 	%f151, 0f00000000;   	// 0
	mov.f32 	%f149, 0f00000000;   	// 0
	mov.f32 	%f147, 0f00000000;   	// 0
	mov.f32 	%f145, 0f00000000;   	// 0
	bra.uni 	$Lt_38_711170;
$Lt_38_711426:
	.loc	20	57	0
	mov.u32 	%r194, 0;
	setp.ne.s32 	%p121, %r41, %r194;
	@%p121 bra 	$Lt_38_478210;
	.loc	20	59	0
	mov.f32 	%f164, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_163;
$Lt_38_478210:
	.loc	20	61	0
	mov.u32 	%r195, 64;
	setp.ne.s32 	%p122, %r41, %r195;
	@%p122 bra 	$Lt_38_478466;
	.loc	20	63	0
	mov.f32 	%f164, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_163;
$Lt_38_478466:
	.loc	20	65	0
	mov.u32 	%r196, 128;
	setp.ne.s32 	%p123, %r41, %r196;
	@%p123 bra 	$Lt_38_478722;
	.loc	20	68	0
	mov.f32 	%f164, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_163;
$Lt_38_478722:
	.loc	20	70	0
	mov.u32 	%r197, 192;
	setp.ne.s32 	%p124, %r41, %r197;
	@%p124 bra 	$Lt_38_478978;
	.loc	20	72	0
	mov.f32 	%f164, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_163;
$Lt_38_478978:
	.loc	20	76	0
	mov.f32 	%f164, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_163:
	.loc	20	447	0
	div.approx.ftz.f32 	%f165, %f164, %f10;
	mul.ftz.f32 	%f147, %f165, %f147;
	.loc	20	448	0
	mul.ftz.f32 	%f149, %f165, %f149;
	.loc	20	449	0
	mul.ftz.f32 	%f151, %f165, %f151;
$Lt_38_711170:
	.loc	20	452	0
	mov.u32 	%r198, 0;
	setp.eq.s32 	%p125, %r46, %r198;
	@%p125 bra 	$Lt_38_711682;
	.loc	20	57	0
	mov.u32 	%r199, 0;
	setp.ne.s32 	%p126, %r41, %r199;
	@%p126 bra 	$Lt_38_479490;
	.loc	20	59	0
	mov.f32 	%f166, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_161;
$Lt_38_479490:
	.loc	20	61	0
	mov.u32 	%r200, 64;
	setp.ne.s32 	%p127, %r41, %r200;
	@%p127 bra 	$Lt_38_479746;
	.loc	20	63	0
	mov.f32 	%f166, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_161;
$Lt_38_479746:
	.loc	20	65	0
	mov.u32 	%r201, 128;
	setp.ne.s32 	%p128, %r41, %r201;
	@%p128 bra 	$Lt_38_480002;
	.loc	20	68	0
	mov.f32 	%f166, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_161;
$Lt_38_480002:
	.loc	20	70	0
	mov.u32 	%r202, 192;
	setp.ne.s32 	%p129, %r41, %r202;
	@%p129 bra 	$Lt_38_480258;
	.loc	20	72	0
	mov.f32 	%f166, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_161;
$Lt_38_480258:
	.loc	20	76	0
	mov.f32 	%f166, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_161:
	.loc	20	107	0
	and.b32 	%r148, %r20, 2048;
	mov.s32 	%r203, 0;
	setp.ne.s32 	%p89, %r148, %r203;
	@!%p89 bra 	$Lt_38_712450;
	.loc	20	100	0
	ld.const.f32 	%f167, [kYCbCrOffset+0];
	bra.uni 	$Lt_38_712194;
$Lt_38_712450:
	ld.const.f32 	%f167, [kYCbCrFullRangeOffset+0];
$Lt_38_712194:
	.loc	20	107	0
	@!%p89 bra 	$Lt_38_712962;
	.loc	20	100	0
	ld.const.f32 	%f168, [kYCbCrOffset+4];
	bra.uni 	$Lt_38_712706;
$Lt_38_712962:
	ld.const.f32 	%f168, [kYCbCrFullRangeOffset+4];
$Lt_38_712706:
	.loc	20	107	0
	@!%p89 bra 	$Lt_38_713474;
	.loc	20	100	0
	ld.const.f32 	%f169, [kYCbCrOffset+8];
	bra.uni 	$Lt_38_713218;
$Lt_38_713474:
	ld.const.f32 	%f169, [kYCbCrFullRangeOffset+8];
$Lt_38_713218:
	.loc	20	454	0
	mov.f32 	%f170, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f171, %f166, %f170;
	fma.rn.ftz.f32 	%f147, %f171, %f167, %f147;
	fma.rn.ftz.f32 	%f149, %f171, %f168, %f149;
	fma.rn.ftz.f32 	%f151, %f171, %f169, %f151;
$Lt_38_711682:
	.loc	20	551	0
	mov.f32 	%f10, %f145;
	mov.f32 	%f12, %f147;
	mov.f32 	%f13, %f149;
	mov.f32 	%f14, %f151;
$Lt_38_704002:
$L_38_679170:
$Lt_38_703490:
	.loc	20	540	0
	and.b32 	%r204, %r20, 4096;
	mov.s32 	%r205, 0;
	setp.eq.s32 	%p130, %r204, %r205;
	@!%p130 bra 	$Lt_38_713730;
	.loc	21	268	0
	mov.f32 	%f172, %f12;
	.loc	21	269	0
	mov.f32 	%f173, %f10;
	.loc	20	558	0
	mov.f32 	%f10, %f14;
	mov.f32 	%f12, %f13;
	mov.f32 	%f13, %f172;
	mov.f32 	%f14, %f173;
$Lt_38_713730:
	@!%p8 bra 	$Lt_38_714242;
	.loc	20	57	0
	mov.u32 	%r206, 0;
	setp.ne.s32 	%p131, %r41, %r206;
	@%p131 bra 	$Lt_38_482818;
	.loc	20	59	0
	mov.f32 	%f174, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_157;
$Lt_38_482818:
	.loc	20	61	0
	mov.u32 	%r207, 64;
	setp.ne.s32 	%p132, %r41, %r207;
	@%p132 bra 	$Lt_38_483074;
	.loc	20	63	0
	mov.f32 	%f174, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_157;
$Lt_38_483074:
	.loc	20	65	0
	mov.u32 	%r208, 128;
	setp.ne.s32 	%p133, %r41, %r208;
	@%p133 bra 	$Lt_38_483330;
	.loc	20	68	0
	mov.f32 	%f174, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_157;
$Lt_38_483330:
	.loc	20	70	0
	mov.u32 	%r209, 192;
	setp.ne.s32 	%p134, %r41, %r209;
	@%p134 bra 	$Lt_38_483586;
	.loc	20	72	0
	mov.f32 	%f174, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_157;
$Lt_38_483586:
	.loc	20	76	0
	mov.f32 	%f174, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_157:
	.loc	20	564	0
	mov.f32 	%f175, 0f3f000000;   	// 0.5
	add.ftz.f32 	%f176, %f10, %f175;
	mov.f32 	%f177, 0f00000000;   	// 0
	mov.f32 	%f178, 0f00000000;   	// 0
	setp.gt.ftz.f32 	%p135, %f176, %f178;
	selp.f32 	%f179, %f176, %f177, %p135;
	min.ftz.f32 	%f10, %f179, %f174;
	mov.f32 	%f180, 0f3f000000;   	// 0.5
	add.ftz.f32 	%f181, %f12, %f180;
	mov.f32 	%f182, 0f00000000;   	// 0
	mov.f32 	%f183, 0f00000000;   	// 0
	setp.gt.ftz.f32 	%p136, %f181, %f183;
	selp.f32 	%f184, %f181, %f182, %p136;
	min.ftz.f32 	%f12, %f184, %f174;
	mov.f32 	%f185, 0f3f000000;   	// 0.5
	add.ftz.f32 	%f186, %f13, %f185;
	mov.f32 	%f187, 0f00000000;   	// 0
	mov.f32 	%f188, 0f00000000;   	// 0
	setp.gt.ftz.f32 	%p137, %f186, %f188;
	selp.f32 	%f189, %f186, %f187, %p137;
	min.ftz.f32 	%f13, %f189, %f174;
	mov.f32 	%f190, 0f3f000000;   	// 0.5
	add.ftz.f32 	%f191, %f14, %f190;
	mov.f32 	%f192, 0f00000000;   	// 0
	mov.f32 	%f193, 0f00000000;   	// 0
	setp.gt.ftz.f32 	%p138, %f191, %f193;
	selp.f32 	%f194, %f191, %f192, %p138;
	min.ftz.f32 	%f14, %f194, %f174;
$Lt_38_714242:
	st.v4.f32 	[%rd2+0], {%f10,%f12,%f13,%f14};
	.loc	22	60	0
	@!%p1 bra 	$Lt_38_715010;
	.loc	20	57	0
	@!%p3 bra 	$Lt_38_485634;
	.loc	20	59	0
	mov.f32 	%f8, 0f437f0000;     	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_155;
$Lt_38_485634:
	.loc	20	61	0
	mov.u32 	%r210, 64;
	setp.ne.s32 	%p139, %r34, %r210;
	@%p139 bra 	$Lt_38_485890;
	.loc	20	63	0
	mov.f32 	%f8, 0f447fc000;     	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_155;
$Lt_38_485890:
	.loc	20	65	0
	mov.u32 	%r211, 128;
	setp.ne.s32 	%p140, %r34, %r211;
	@%p140 bra 	$Lt_38_486146;
	.loc	20	68	0
	mov.f32 	%f8, 0f47000000;     	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_155;
$Lt_38_486146:
	.loc	20	70	0
	mov.u32 	%r212, 192;
	setp.ne.s32 	%p141, %r34, %r212;
	@%p141 bra 	$Lt_38_486402;
	.loc	20	72	0
	mov.f32 	%f8, 0fbf800000;     	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_155;
$Lt_38_486402:
	.loc	20	76	0
	mov.f32 	%f8, 0f3f800000;     	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_155:
	.loc	20	469	0
	cvt.u8.u32 	%r213, %r16;
	cvt.rn.f32.u32 	%f195, %r213;
	add.ftz.f32 	%f196, %f195, %f7;
	cvt.rzi.ftz.u32.f32 	%r214, %f196;
	shr.u32 	%r215, %r214, 1;
	cvt.rn.f32.u32 	%f197, %r215;
	mov.f32 	%f10, %f197;
	cvt.u8.u32 	%r216, %r12;
	cvt.rn.f32.u32 	%f198, %r216;
	add.ftz.f32 	%f199, %f198, %f3;
	cvt.rzi.ftz.u32.f32 	%r217, %f199;
	shr.u32 	%r218, %r217, 1;
	cvt.rn.f32.u32 	%f200, %r218;
	mov.f32 	%f12, %f200;
	mov.f32 	%f13, %f5;
	mov.f32 	%f14, %f8;
	@!%p7 bra 	$Lt_38_715266;
	.loc	20	473	0
	mov.f32 	%f10, %f14;
	mov.f32 	%f12, %f5;
	mov.f32 	%f13, %f200;
	mov.f32 	%f14, %f197;
$Lt_38_715266:
	.loc	20	476	0
	mov.u32 	%r219, 0;
	setp.eq.s32 	%p142, %r51, %r219;
	@%p142 bra 	$Lt_38_774658;
	mov.u32 	%r220, 0;
	setp.ne.s32 	%p143, %r46, %r220;
	@%p143 bra 	$Lt_38_774658;
	.loc	20	57	0
	@!%p3 bra 	$Lt_38_486914;
	.loc	20	59	0
	mov.f32 	%f15, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_153;
$Lt_38_486914:
	.loc	20	61	0
	mov.u32 	%r221, 64;
	setp.ne.s32 	%p144, %r34, %r221;
	@%p144 bra 	$Lt_38_487170;
	.loc	20	63	0
	mov.f32 	%f15, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_153;
$Lt_38_487170:
	.loc	20	65	0
	mov.u32 	%r222, 128;
	setp.ne.s32 	%p145, %r34, %r222;
	@%p145 bra 	$Lt_38_487426;
	.loc	20	68	0
	mov.f32 	%f15, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_153;
$Lt_38_487426:
	.loc	20	70	0
	mov.u32 	%r223, 192;
	setp.ne.s32 	%p146, %r34, %r223;
	@%p146 bra 	$Lt_38_487682;
	.loc	20	72	0
	mov.f32 	%f15, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_153;
$Lt_38_487682:
	.loc	20	76	0
	mov.f32 	%f15, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_153:
	.loc	20	118	0
	and.b32 	%r57, %r33, 2048;
	mov.s32 	%r224, 0;
	setp.ne.s32 	%p15, %r57, %r224;
	@!%p15 bra 	$Lt_38_716034;
	.loc	20	100	0
	ld.const.f32 	%f201, [kYCbCrOffset+0];
	bra.uni 	$Lt_38_715778;
$Lt_38_716034:
	ld.const.f32 	%f201, [kYCbCrFullRangeOffset+0];
$Lt_38_715778:
	.loc	20	118	0
	@!%p15 bra 	$Lt_38_716546;
	.loc	20	100	0
	ld.const.f32 	%f202, [kYCbCrOffset+4];
	bra.uni 	$Lt_38_716290;
$Lt_38_716546:
	ld.const.f32 	%f202, [kYCbCrFullRangeOffset+4];
$Lt_38_716290:
	.loc	20	118	0
	@!%p15 bra 	$Lt_38_717058;
	.loc	20	100	0
	ld.const.f32 	%f203, [kYCbCrOffset+8];
	bra.uni 	$Lt_38_716802;
$Lt_38_717058:
	ld.const.f32 	%f203, [kYCbCrFullRangeOffset+8];
$Lt_38_716802:
	.loc	20	478	0
	mov.f32 	%f204, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f205, %f15, %f204;
	mul.ftz.f32 	%f206, %f205, %f201;
	sub.ftz.f32 	%f12, %f12, %f206;
	mul.ftz.f32 	%f207, %f205, %f202;
	sub.ftz.f32 	%f13, %f13, %f207;
	mul.ftz.f32 	%f208, %f205, %f203;
	sub.ftz.f32 	%f14, %f14, %f208;
$Lt_38_774658:
$Lt_38_193538:
	.loc	20	481	0
	@!%p16 bra 	$Lt_38_775170;
	mov.u32 	%r225, 0;
	setp.ne.s32 	%p147, %r59, %r225;
	@%p147 bra 	$Lt_38_775170;
	.loc	20	483	0
	mov.f32 	%f209, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p148, %f12, %f209;
	@!%p148 bra 	$Lt_38_717314;
	.loc	20	372	0
	neg.ftz.f32 	%f210, %f12;
	lg2.approx.ftz.f32 	%f211, %f210;
	mov.f32 	%f212, 0f3ee66666;   	// 0.45
	mul.ftz.f32 	%f213, %f211, %f212;
	ex2.approx.ftz.f32 	%f214, %f213;
	neg.ftz.f32 	%f30, %f214;
	bra.uni 	$LDWendi___log2f_216_149;
$Lt_38_717314:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f215, %f12;
	mov.f32 	%f216, 0f3ee66666;   	// 0.45
	mul.ftz.f32 	%f217, %f215, %f216;
	ex2.approx.ftz.f32 	%f30, %f217;
$LDWendi___log2f_216_149:
	.loc	20	483	0
	mov.f32 	%f218, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p149, %f13, %f218;
	@!%p149 bra 	$Lt_38_717826;
	.loc	20	372	0
	neg.ftz.f32 	%f219, %f13;
	lg2.approx.ftz.f32 	%f220, %f219;
	mov.f32 	%f221, 0f3ee66666;   	// 0.45
	mul.ftz.f32 	%f222, %f220, %f221;
	ex2.approx.ftz.f32 	%f223, %f222;
	neg.ftz.f32 	%f40, %f223;
	bra.uni 	$LDWendi___log2f_216_147;
$Lt_38_717826:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f224, %f13;
	mov.f32 	%f225, 0f3ee66666;   	// 0.45
	mul.ftz.f32 	%f226, %f224, %f225;
	ex2.approx.ftz.f32 	%f40, %f226;
$LDWendi___log2f_216_147:
	.loc	20	483	0
	mov.f32 	%f227, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p150, %f14, %f227;
	@!%p150 bra 	$Lt_38_718338;
	.loc	20	372	0
	neg.ftz.f32 	%f228, %f14;
	lg2.approx.ftz.f32 	%f229, %f228;
	mov.f32 	%f230, 0f3ee66666;   	// 0.45
	mul.ftz.f32 	%f231, %f229, %f230;
	ex2.approx.ftz.f32 	%f232, %f231;
	neg.ftz.f32 	%f50, %f232;
	bra.uni 	$LDWendi___log2f_216_145;
$Lt_38_718338:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f233, %f14;
	mov.f32 	%f234, 0f3ee66666;   	// 0.45
	mul.ftz.f32 	%f235, %f233, %f234;
	ex2.approx.ftz.f32 	%f50, %f235;
$LDWendi___log2f_216_145:
	.loc	20	483	0
	mov.f32 	%f12, %f30;
	mov.f32 	%f13, %f40;
	mov.f32 	%f14, %f50;
$Lt_38_775170:
$Lt_38_197378:
	.loc	20	486	0
	mov.u32 	%r226, 0;
	setp.ne.s32 	%p151, %r69, %r226;
	@%p151 bra 	$Lt_38_248066;
	setp.eq.s32 	%p152, %r46, %r51;
	@%p152 bra 	$Lt_38_248322;
$Lt_38_248066:
	.loc	20	490	0
	mov.u32 	%r227, 0;
	setp.ne.s32 	%p153, %r65, %r227;
	@%p153 bra 	$Lt_38_719106;
	mov.s32 	%r228, 256;
	setp.eq.s32 	%p26, %r41, %r228;
	mov.u32 	%r229, 256;
	setp.ne.s32 	%p154, %r34, %r229;
	@%p154 bra 	$Lt_38_719618;
	.loc	20	137	0
	mov.s32 	%r230, 512;
	setp.eq.s32 	%p28, %r64, %r230;
	@!%p26 bra 	$Lt_38_491010;
	.loc	20	139	0
	@!%p28 bra 	$Lt_38_491266;
	.loc	20	141	0
	mov.s64 	%rd6, %rd5;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_143;
$Lt_38_491266:
	.loc	20	145	0
	mov.s64 	%rd6, %rd7;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_143;
$Lt_38_491010:
	.loc	20	150	0
	@!%p28 bra 	$Lt_38_491522;
	.loc	20	152	0
	mov.s64 	%rd6, %rd8;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_143;
$Lt_38_491522:
	.loc	20	154	0
	and.b32 	%r231, %r20, 2048;
	mov.u32 	%r232, 0;
	setp.ne.s32 	%p155, %r231, %r232;
	@%p155 bra 	$Lt_38_491778;
	.loc	20	156	0
	mov.s64 	%rd6, %rd9;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_143;
$Lt_38_491778:
	.loc	20	160	0
	mov.s64 	%rd6, %rd10;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_143;
$Lt_38_719618:
	@!%p26 bra 	$Lt_38_720130;
	bra.uni 	$Lt_38_718850;
$Lt_38_720130:
	.loc	20	179	0
	mov.u32 	%r233, 512;
	setp.ne.s32 	%p156, %r64, %r233;
	@%p156 bra 	$Lt_38_492546;
	.loc	20	181	0
	mov.s64 	%rd6, %rd11;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_143;
$Lt_38_492546:
	.loc	20	183	0
	and.b32 	%r234, %r20, 2048;
	mov.u32 	%r235, 0;
	setp.ne.s32 	%p157, %r234, %r235;
	@%p157 bra 	$Lt_38_492802;
	.loc	20	185	0
	mov.s64 	%rd6, %rd12;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_143;
$Lt_38_492802:
	.loc	20	189	0
	mov.s64 	%rd6, %rd13;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_143;
$Lt_38_719106:
	mov.s32 	%r236, 0;
	setp.eq.s32 	%p32, %r63, %r236;
	mov.u32 	%r237, 512;
	setp.ne.s32 	%p158, %r66, %r237;
	@%p158 bra 	$Lt_38_720642;
	mov.s32 	%r238, 256;
	setp.eq.s32 	%p34, %r34, %r238;
	@!%p32 bra 	$Lt_38_721154;
	mov.s32 	%r239, 256;
	setp.eq.s32 	%p26, %r41, %r239;
	@!%p34 bra 	$Lt_38_721666;
	@!%p26 bra 	$Lt_38_718850;
	.loc	20	202	0
	mov.s64 	%rd6, %rd14;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_143;
$Lt_38_721666:
	.loc	20	211	0
	@!%p26 bra 	$Lt_38_494082;
	.loc	20	213	0
	mov.s64 	%rd6, %rd15;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_143;
$Lt_38_494082:
	.loc	20	217	0
	mov.s64 	%rd6, %rd16;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_143;
$Lt_38_721154:
	@!%p34 bra 	$Lt_38_722690;
	bra.uni 	$Lt_38_718850;
$Lt_38_722690:
	mov.s32 	%r240, 256;
	set.eq.u32.s32 	%r241, %r41, %r240;
	neg.s32 	%r242, %r241;
	and.b32 	%r243, %r20, 2048;
	mov.s32 	%r244, 0;
	set.eq.u32.s32 	%r245, %r243, %r244;
	neg.s32 	%r246, %r245;
	or.b32 	%r247, %r242, %r246;
	mov.u32 	%r248, 0;
	setp.eq.s32 	%p159, %r247, %r248;
	@%p159 bra 	$Lt_38_723202;
	bra.uni 	$Lt_38_718850;
$Lt_38_723202:
	.loc	20	250	0
	mov.s64 	%rd6, %rd17;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_143;
$Lt_38_720642:
	and.b32 	%r57, %r33, 2048;
	mov.s32 	%r249, 0;
	setp.eq.s32 	%p160, %r57, %r249;
	mov.pred 	%p37, %p160;
	mov.pred 	%p161, %p39;
	@!%p160 bra 	$Lt_38_723714;
	@!%p32 bra 	$Lt_38_718850;
	.loc	20	259	0
	mov.u32 	%r250, 256;
	setp.ne.s32 	%p162, %r41, %r250;
	@%p162 bra 	$Lt_38_495618;
	.loc	20	261	0
	mov.s64 	%rd6, %rd18;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_143;
$Lt_38_495618:
	.loc	20	265	0
	mov.s64 	%rd6, %rd19;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_143;
$Lt_38_723714:
	mov.s32 	%r251, 256;
	setp.eq.s32 	%p34, %r34, %r251;
	@!%p32 bra 	$Lt_38_724738;
	mov.s32 	%r252, 256;
	setp.eq.s32 	%p26, %r41, %r252;
	@!%p34 bra 	$Lt_38_725250;
	@!%p26 bra 	$Lt_38_718850;
	.loc	20	302	0
	mov.s64 	%rd6, %rd20;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_143;
$Lt_38_725250:
	.loc	20	311	0
	@!%p26 bra 	$Lt_38_497410;
	.loc	20	313	0
	mov.s64 	%rd6, %rd21;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_143;
$Lt_38_497410:
	.loc	20	317	0
	mov.s64 	%rd6, %rd22;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_143;
$Lt_38_724738:
	@!%p34 bra 	$Lt_38_726274;
	bra.uni 	$Lt_38_718850;
$Lt_38_726274:
	selp.s32 	%r253, 1, 0, %p37;
	mov.s32 	%r254, 256;
	set.eq.u32.s32 	%r255, %r41, %r254;
	neg.s32 	%r256, %r255;
	or.b32 	%r257, %r253, %r256;
	mov.u32 	%r258, 0;
	setp.eq.s32 	%p163, %r257, %r258;
	@%p163 bra 	$Lt_38_726786;
	bra.uni 	$Lt_38_718850;
$Lt_38_726786:
	.loc	20	350	0
	mov.s64 	%rd6, %rd23;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_143;
$Lt_38_718850:
	.loc	20	355	0
	mov.u64 	%rd6, 0;
$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_143:
	.loc	20	490	0
	ld.f32 	%f236, [%rd6+16];
	mul.ftz.f32 	%f237, %f236, %f13;
	ld.f32 	%f238, [%rd6+12];
	fma.rn.ftz.f32 	%f239, %f238, %f12, %f237;
	ld.f32 	%f240, [%rd6+20];
	fma.rn.ftz.f32 	%f241, %f240, %f14, %f239;
	ld.f32 	%f242, [%rd6+28];
	mul.ftz.f32 	%f243, %f242, %f13;
	ld.f32 	%f244, [%rd6+24];
	fma.rn.ftz.f32 	%f245, %f244, %f12, %f243;
	ld.f32 	%f246, [%rd6+32];
	fma.rn.ftz.f32 	%f247, %f246, %f14, %f245;
	ld.f32 	%f248, [%rd6+4];
	mul.ftz.f32 	%f249, %f248, %f13;
	ld.f32 	%f250, [%rd6+0];
	fma.rn.ftz.f32 	%f251, %f250, %f12, %f249;
	ld.f32 	%f252, [%rd6+8];
	fma.rn.ftz.f32 	%f12, %f252, %f14, %f251;
	mov.f32 	%f13, %f241;
	mov.f32 	%f14, %f247;
	@!%p42 bra 	$Lt_38_727554;
	.loc	20	494	0
	mov.s32 	%r259, 256;
	setp.eq.s32 	%p34, %r34, %r259;
	@!%p34 bra 	$L_38_681218;
	mov.s32 	%r260, 0;
	setp.eq.s32 	%p43, %r41, %r260;
	@%p43 bra 	$L_38_680962;
$L_38_681218:
	@!%p3 bra 	$Lt_38_776450;
	mov.u32 	%r261, 256;
	setp.ne.s32 	%p164, %r41, %r261;
	@%p164 bra 	$Lt_38_776450;
	mov.s32 	%r262, 0;
	setp.eq.s32 	%p43, %r41, %r262;
$L_38_680962:
	.loc	20	57	0
	@!%p43 bra 	$Lt_38_498946;
	.loc	20	59	0
	mov.f32 	%f71, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_141;
$Lt_38_498946:
	.loc	20	61	0
	mov.u32 	%r263, 64;
	setp.ne.s32 	%p165, %r41, %r263;
	@%p165 bra 	$Lt_38_499202;
	.loc	20	63	0
	mov.f32 	%f71, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_141;
$Lt_38_499202:
	.loc	20	65	0
	mov.u32 	%r264, 128;
	setp.ne.s32 	%p166, %r41, %r264;
	@%p166 bra 	$Lt_38_499458;
	.loc	20	68	0
	mov.f32 	%f71, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_141;
$Lt_38_499458:
	.loc	20	70	0
	mov.u32 	%r265, 192;
	setp.ne.s32 	%p167, %r41, %r265;
	@%p167 bra 	$Lt_38_499714;
	.loc	20	72	0
	mov.f32 	%f71, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_141;
$Lt_38_499714:
	.loc	20	76	0
	mov.f32 	%f71, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_141:
	.loc	20	57	0
	@!%p3 bra 	$Lt_38_499970;
	.loc	20	59	0
	mov.f32 	%f72, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_139;
$Lt_38_499970:
	.loc	20	61	0
	mov.u32 	%r266, 64;
	setp.ne.s32 	%p168, %r34, %r266;
	@%p168 bra 	$Lt_38_500226;
	.loc	20	63	0
	mov.f32 	%f72, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_139;
$Lt_38_500226:
	.loc	20	65	0
	mov.u32 	%r267, 128;
	setp.ne.s32 	%p169, %r34, %r267;
	@%p169 bra 	$Lt_38_500482;
	.loc	20	68	0
	mov.f32 	%f72, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_139;
$Lt_38_500482:
	.loc	20	70	0
	mov.u32 	%r268, 192;
	setp.ne.s32 	%p170, %r34, %r268;
	@%p170 bra 	$Lt_38_500738;
	.loc	20	72	0
	mov.f32 	%f72, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_139;
$Lt_38_500738:
	.loc	20	76	0
	mov.f32 	%f72, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_139:
	.loc	20	498	0
	div.approx.ftz.f32 	%f253, %f71, %f72;
	mul.ftz.f32 	%f10, %f10, %f253;
	bra.uni 	$Lt_38_727554;
$Lt_38_776450:
$L_38_680706:
	.loc	20	500	0
	@!%p34 bra 	$L_38_682242;
	@%p8 bra 	$L_38_681986;
$L_38_682242:
	@!%p9 bra 	$Lt_38_777474;
	mov.u32 	%r269, 256;
	setp.ne.s32 	%p171, %r41, %r269;
	@%p171 bra 	$Lt_38_777474;
$L_38_681986:
	.loc	20	57	0
	mov.u32 	%r270, 0;
	setp.ne.s32 	%p172, %r41, %r270;
	@%p172 bra 	$Lt_38_501250;
	.loc	20	59	0
	mov.f32 	%f71, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_137;
$Lt_38_501250:
	.loc	20	61	0
	mov.u32 	%r271, 64;
	setp.ne.s32 	%p173, %r41, %r271;
	@%p173 bra 	$Lt_38_501506;
	.loc	20	63	0
	mov.f32 	%f71, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_137;
$Lt_38_501506:
	.loc	20	65	0
	mov.u32 	%r272, 128;
	setp.ne.s32 	%p174, %r41, %r272;
	@%p174 bra 	$Lt_38_501762;
	.loc	20	68	0
	mov.f32 	%f71, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_137;
$Lt_38_501762:
	.loc	20	70	0
	mov.u32 	%r273, 192;
	setp.ne.s32 	%p175, %r41, %r273;
	@%p175 bra 	$Lt_38_502018;
	.loc	20	72	0
	mov.f32 	%f71, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_137;
$Lt_38_502018:
	.loc	20	76	0
	mov.f32 	%f71, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_137:
	.loc	20	57	0
	@!%p3 bra 	$Lt_38_502274;
	.loc	20	59	0
	mov.f32 	%f72, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_135;
$Lt_38_502274:
	.loc	20	61	0
	mov.u32 	%r274, 64;
	setp.ne.s32 	%p176, %r34, %r274;
	@%p176 bra 	$Lt_38_502530;
	.loc	20	63	0
	mov.f32 	%f72, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_135;
$Lt_38_502530:
	.loc	20	65	0
	mov.u32 	%r275, 128;
	setp.ne.s32 	%p177, %r34, %r275;
	@%p177 bra 	$Lt_38_502786;
	.loc	20	68	0
	mov.f32 	%f72, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_135;
$Lt_38_502786:
	.loc	20	70	0
	mov.u32 	%r276, 192;
	setp.ne.s32 	%p178, %r34, %r276;
	@%p178 bra 	$Lt_38_503042;
	.loc	20	72	0
	mov.f32 	%f72, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_135;
$Lt_38_503042:
	.loc	20	76	0
	mov.f32 	%f72, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_135:
	.loc	20	504	0
	div.approx.ftz.f32 	%f254, %f71, %f72;
	mul.ftz.f32 	%f10, %f10, %f254;
	.loc	20	57	0
	@!%p3 bra 	$Lt_38_504322;
	.loc	20	59	0
	mov.f32 	%f72, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_131;
$Lt_38_504322:
	.loc	20	61	0
	mov.u32 	%r277, 64;
	setp.ne.s32 	%p179, %r34, %r277;
	@%p179 bra 	$Lt_38_504578;
	.loc	20	63	0
	mov.f32 	%f72, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_131;
$Lt_38_504578:
	.loc	20	65	0
	mov.u32 	%r278, 128;
	setp.ne.s32 	%p180, %r34, %r278;
	@%p180 bra 	$Lt_38_504834;
	.loc	20	68	0
	mov.f32 	%f72, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_131;
$Lt_38_504834:
	.loc	20	70	0
	mov.u32 	%r279, 192;
	setp.ne.s32 	%p181, %r34, %r279;
	@%p181 bra 	$Lt_38_505090;
	.loc	20	72	0
	mov.f32 	%f72, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_131;
$Lt_38_505090:
	.loc	20	76	0
	mov.f32 	%f72, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_131:
	.loc	20	505	0
	mov.f32 	%f255, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f256, %f255, %f72;
	mul.ftz.f32 	%f12, %f12, %f256;
	.loc	20	57	0
	@!%p3 bra 	$Lt_38_506370;
	.loc	20	59	0
	mov.f32 	%f72, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_127;
$Lt_38_506370:
	.loc	20	61	0
	mov.u32 	%r280, 64;
	setp.ne.s32 	%p182, %r34, %r280;
	@%p182 bra 	$Lt_38_506626;
	.loc	20	63	0
	mov.f32 	%f72, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_127;
$Lt_38_506626:
	.loc	20	65	0
	mov.u32 	%r281, 128;
	setp.ne.s32 	%p183, %r34, %r281;
	@%p183 bra 	$Lt_38_506882;
	.loc	20	68	0
	mov.f32 	%f72, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_127;
$Lt_38_506882:
	.loc	20	70	0
	mov.u32 	%r282, 192;
	setp.ne.s32 	%p184, %r34, %r282;
	@%p184 bra 	$Lt_38_507138;
	.loc	20	72	0
	mov.f32 	%f72, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_127;
$Lt_38_507138:
	.loc	20	76	0
	mov.f32 	%f72, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_127:
	.loc	20	506	0
	mov.f32 	%f257, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f258, %f257, %f72;
	mul.ftz.f32 	%f13, %f241, %f258;
	.loc	20	57	0
	@!%p3 bra 	$Lt_38_508418;
	.loc	20	59	0
	mov.f32 	%f72, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_123;
$Lt_38_508418:
	.loc	20	61	0
	mov.u32 	%r283, 64;
	setp.ne.s32 	%p185, %r34, %r283;
	@%p185 bra 	$Lt_38_508674;
	.loc	20	63	0
	mov.f32 	%f72, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_123;
$Lt_38_508674:
	.loc	20	65	0
	mov.u32 	%r284, 128;
	setp.ne.s32 	%p186, %r34, %r284;
	@%p186 bra 	$Lt_38_508930;
	.loc	20	68	0
	mov.f32 	%f72, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_123;
$Lt_38_508930:
	.loc	20	70	0
	mov.u32 	%r285, 192;
	setp.ne.s32 	%p187, %r34, %r285;
	@%p187 bra 	$Lt_38_509186;
	.loc	20	72	0
	mov.f32 	%f72, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_123;
$Lt_38_509186:
	.loc	20	76	0
	mov.f32 	%f72, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_123:
	.loc	20	507	0
	mov.f32 	%f259, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f260, %f259, %f72;
	mul.ftz.f32 	%f14, %f247, %f260;
	bra.uni 	$Lt_38_727554;
$Lt_38_777474:
$L_38_681730:
	.loc	20	57	0
	@!%p3 bra 	$Lt_38_509442;
	.loc	20	59	0
	mov.f32 	%f71, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_121;
$Lt_38_509442:
	.loc	20	61	0
	mov.u32 	%r286, 64;
	setp.ne.s32 	%p188, %r34, %r286;
	@%p188 bra 	$Lt_38_509698;
	.loc	20	63	0
	mov.f32 	%f71, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_121;
$Lt_38_509698:
	.loc	20	65	0
	mov.u32 	%r287, 128;
	setp.ne.s32 	%p189, %r34, %r287;
	@%p189 bra 	$Lt_38_509954;
	.loc	20	68	0
	mov.f32 	%f71, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_121;
$Lt_38_509954:
	.loc	20	70	0
	mov.u32 	%r288, 192;
	setp.ne.s32 	%p190, %r34, %r288;
	@%p190 bra 	$Lt_38_510210;
	.loc	20	72	0
	mov.f32 	%f71, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_121;
$Lt_38_510210:
	.loc	20	76	0
	mov.f32 	%f71, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_121:
	.loc	20	511	0
	mov.f32 	%f261, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f262, %f71, %f261;
	mul.ftz.f32 	%f10, %f262, %f10;
	mul.ftz.f32 	%f12, %f262, %f12;
	mul.ftz.f32 	%f13, %f262, %f241;
	mul.ftz.f32 	%f14, %f262, %f247;
	bra.uni 	$Lt_38_727554;
$Lt_38_248322:
	.loc	20	486	0
	@!%p42 bra 	$Lt_38_727554;
	.loc	20	57	0
	mov.u32 	%r289, 0;
	setp.ne.s32 	%p191, %r41, %r289;
	@%p191 bra 	$Lt_38_511746;
	.loc	20	59	0
	mov.f32 	%f71, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_117;
$Lt_38_511746:
	.loc	20	61	0
	mov.u32 	%r290, 64;
	setp.ne.s32 	%p192, %r41, %r290;
	@%p192 bra 	$Lt_38_512002;
	.loc	20	63	0
	mov.f32 	%f71, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_117;
$Lt_38_512002:
	.loc	20	65	0
	mov.u32 	%r291, 128;
	setp.ne.s32 	%p193, %r41, %r291;
	@%p193 bra 	$Lt_38_512258;
	.loc	20	68	0
	mov.f32 	%f71, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_117;
$Lt_38_512258:
	.loc	20	70	0
	mov.u32 	%r292, 192;
	setp.ne.s32 	%p194, %r41, %r292;
	@%p194 bra 	$Lt_38_512514;
	.loc	20	72	0
	mov.f32 	%f71, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_117;
$Lt_38_512514:
	.loc	20	76	0
	mov.f32 	%f71, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_117:
	.loc	20	57	0
	@!%p3 bra 	$Lt_38_512770;
	.loc	20	59	0
	mov.f32 	%f72, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_115;
$Lt_38_512770:
	.loc	20	61	0
	mov.u32 	%r293, 64;
	setp.ne.s32 	%p195, %r34, %r293;
	@%p195 bra 	$Lt_38_513026;
	.loc	20	63	0
	mov.f32 	%f72, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_115;
$Lt_38_513026:
	.loc	20	65	0
	mov.u32 	%r294, 128;
	setp.ne.s32 	%p196, %r34, %r294;
	@%p196 bra 	$Lt_38_513282;
	.loc	20	68	0
	mov.f32 	%f72, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_115;
$Lt_38_513282:
	.loc	20	70	0
	mov.u32 	%r295, 192;
	setp.ne.s32 	%p197, %r34, %r295;
	@%p197 bra 	$Lt_38_513538;
	.loc	20	72	0
	mov.f32 	%f72, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_115;
$Lt_38_513538:
	.loc	20	76	0
	mov.f32 	%f72, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_115:
	.loc	20	517	0
	div.approx.ftz.f32 	%f263, %f71, %f72;
	mul.ftz.f32 	%f10, %f263, %f10;
	mul.ftz.f32 	%f12, %f263, %f12;
	mul.ftz.f32 	%f13, %f263, %f13;
	mul.ftz.f32 	%f14, %f263, %f14;
$Lt_38_727554:
$Lt_38_250626:
	.loc	20	520	0
	@!%p78 bra 	$Lt_38_777986;
	mov.u32 	%r296, 0;
	setp.ne.s32 	%p198, %r60, %r296;
	@%p198 bra 	$Lt_38_777986;
	.loc	20	522	0
	mov.f32 	%f264, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p199, %f12, %f264;
	@!%p199 bra 	$Lt_38_728066;
	.loc	20	372	0
	neg.ftz.f32 	%f265, %f12;
	lg2.approx.ftz.f32 	%f266, %f265;
	mov.f32 	%f267, 0f400e38e4;   	// 2.22222
	mul.ftz.f32 	%f268, %f266, %f267;
	ex2.approx.ftz.f32 	%f269, %f268;
	neg.ftz.f32 	%f90, %f269;
	bra.uni 	$LDWendi___log2f_216_113;
$Lt_38_728066:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f270, %f12;
	mov.f32 	%f271, 0f400e38e4;   	// 2.22222
	mul.ftz.f32 	%f272, %f270, %f271;
	ex2.approx.ftz.f32 	%f90, %f272;
$LDWendi___log2f_216_113:
	.loc	20	522	0
	mov.f32 	%f273, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p200, %f13, %f273;
	@!%p200 bra 	$Lt_38_728578;
	.loc	20	372	0
	neg.ftz.f32 	%f274, %f13;
	lg2.approx.ftz.f32 	%f275, %f274;
	mov.f32 	%f276, 0f400e38e4;   	// 2.22222
	mul.ftz.f32 	%f277, %f275, %f276;
	ex2.approx.ftz.f32 	%f278, %f277;
	neg.ftz.f32 	%f100, %f278;
	bra.uni 	$LDWendi___log2f_216_111;
$Lt_38_728578:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f279, %f13;
	mov.f32 	%f280, 0f400e38e4;   	// 2.22222
	mul.ftz.f32 	%f281, %f279, %f280;
	ex2.approx.ftz.f32 	%f100, %f281;
$LDWendi___log2f_216_111:
	.loc	20	522	0
	mov.f32 	%f282, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p201, %f14, %f282;
	@!%p201 bra 	$Lt_38_729090;
	.loc	20	372	0
	neg.ftz.f32 	%f283, %f14;
	lg2.approx.ftz.f32 	%f284, %f283;
	mov.f32 	%f285, 0f400e38e4;   	// 2.22222
	mul.ftz.f32 	%f286, %f284, %f285;
	ex2.approx.ftz.f32 	%f287, %f286;
	neg.ftz.f32 	%f110, %f287;
	bra.uni 	$LDWendi___log2f_216_109;
$Lt_38_729090:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f288, %f14;
	mov.f32 	%f289, 0f400e38e4;   	// 2.22222
	mul.ftz.f32 	%f290, %f288, %f289;
	ex2.approx.ftz.f32 	%f110, %f290;
$LDWendi___log2f_216_109:
	.loc	20	522	0
	mov.f32 	%f12, %f90;
	mov.f32 	%f13, %f100;
	mov.f32 	%f14, %f110;
$Lt_38_777986:
$Lt_38_252674:
	.loc	20	525	0
	mov.u32 	%r297, 0;
	setp.eq.s32 	%p202, %r46, %r297;
	@%p202 bra 	$Lt_38_778498;
	mov.u32 	%r298, 0;
	setp.ne.s32 	%p203, %r51, %r298;
	@%p203 bra 	$Lt_38_778498;
	.loc	20	57	0
	mov.u32 	%r299, 0;
	setp.ne.s32 	%p204, %r41, %r299;
	@%p204 bra 	$Lt_38_514562;
	.loc	20	59	0
	mov.f32 	%f114, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_107;
$Lt_38_514562:
	.loc	20	61	0
	mov.u32 	%r300, 64;
	setp.ne.s32 	%p205, %r41, %r300;
	@%p205 bra 	$Lt_38_514818;
	.loc	20	63	0
	mov.f32 	%f114, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_107;
$Lt_38_514818:
	.loc	20	65	0
	mov.u32 	%r301, 128;
	setp.ne.s32 	%p206, %r41, %r301;
	@%p206 bra 	$Lt_38_515074;
	.loc	20	68	0
	mov.f32 	%f114, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_107;
$Lt_38_515074:
	.loc	20	70	0
	mov.u32 	%r302, 192;
	setp.ne.s32 	%p207, %r41, %r302;
	@%p207 bra 	$Lt_38_515330;
	.loc	20	72	0
	mov.f32 	%f114, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_107;
$Lt_38_515330:
	.loc	20	76	0
	mov.f32 	%f114, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_107:
	.loc	20	107	0
	and.b32 	%r148, %r20, 2048;
	mov.s32 	%r303, 0;
	setp.ne.s32 	%p89, %r148, %r303;
	@!%p89 bra 	$Lt_38_729858;
	.loc	20	100	0
	ld.const.f32 	%f291, [kYCbCrOffset+0];
	bra.uni 	$Lt_38_729602;
$Lt_38_729858:
	ld.const.f32 	%f291, [kYCbCrFullRangeOffset+0];
$Lt_38_729602:
	.loc	20	107	0
	@!%p89 bra 	$Lt_38_730370;
	.loc	20	100	0
	ld.const.f32 	%f292, [kYCbCrOffset+4];
	bra.uni 	$Lt_38_730114;
$Lt_38_730370:
	ld.const.f32 	%f292, [kYCbCrFullRangeOffset+4];
$Lt_38_730114:
	.loc	20	107	0
	@!%p89 bra 	$Lt_38_730882;
	.loc	20	100	0
	ld.const.f32 	%f293, [kYCbCrOffset+8];
	bra.uni 	$Lt_38_730626;
$Lt_38_730882:
	ld.const.f32 	%f293, [kYCbCrFullRangeOffset+8];
$Lt_38_730626:
	.loc	20	527	0
	mov.f32 	%f294, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f295, %f114, %f294;
	fma.rn.ftz.f32 	%f12, %f295, %f291, %f12;
	fma.rn.ftz.f32 	%f13, %f295, %f292, %f13;
	fma.rn.ftz.f32 	%f14, %f295, %f293, %f14;
$Lt_38_778498:
$Lt_38_259074:
	.loc	20	525	0
	@!%p90 bra 	$Lt_38_731650;
	.loc	20	532	0
	mov.u32 	%r304, 8;
	setp.ne.s32 	%p208, %r151, %r304;
	@%p208 bra 	$L_38_683522;
	mov.u32 	%r305, 12;
	setp.eq.s32 	%p209, %r150, %r305;
	@%p209 bra 	$Lt_38_779266;
$L_38_683522:
	mov.u32 	%r306, 12;
	setp.eq.s32 	%p210, %r151, %r306;
	@%p210 bra 	$Lt_38_779266;
	mov.u32 	%r307, 0;
	setp.ne.s32 	%p211, %r151, %r307;
	@%p211 bra 	$L_38_682754;
$Lt_38_779266:
$L_38_683010:
	.loc	20	57	0
	mov.u32 	%r308, 0;
	setp.ne.s32 	%p212, %r41, %r308;
	@%p212 bra 	$Lt_38_517890;
	.loc	20	59	0
	mov.f32 	%f120, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_103;
$Lt_38_517890:
	.loc	20	61	0
	mov.u32 	%r309, 64;
	setp.ne.s32 	%p213, %r41, %r309;
	@%p213 bra 	$Lt_38_518146;
	.loc	20	63	0
	mov.f32 	%f120, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_103;
$Lt_38_518146:
	.loc	20	65	0
	mov.u32 	%r310, 128;
	setp.ne.s32 	%p214, %r41, %r310;
	@%p214 bra 	$Lt_38_518402;
	.loc	20	68	0
	mov.f32 	%f120, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_103;
$Lt_38_518402:
	.loc	20	70	0
	mov.u32 	%r311, 192;
	setp.ne.s32 	%p215, %r41, %r311;
	@%p215 bra 	$Lt_38_518658;
	.loc	20	72	0
	mov.f32 	%f120, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_103;
$Lt_38_518658:
	.loc	20	76	0
	mov.f32 	%f120, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_103:
	.loc	20	536	0
	mov.f32 	%f10, %f120;
	bra.uni 	$Lt_38_731650;
$L_38_682754:
	.loc	20	540	0
	mov.s32 	%r312, 12;
	setp.eq.s32 	%p99, %r150, %r312;
	mov.s32 	%r313, 4;
	set.eq.u32.s32 	%r314, %r151, %r313;
	neg.s32 	%r315, %r314;
	selp.s32 	%r316, 1, 0, %p99;
	mov.s32 	%r317, 8;
	set.eq.u32.s32 	%r318, %r150, %r317;
	neg.s32 	%r319, %r318;
	or.b32 	%r320, %r316, %r319;
	and.b32 	%r321, %r315, %r320;
	mov.u32 	%r322, 0;
	setp.eq.s32 	%p216, %r321, %r322;
	@%p216 bra 	$Lt_38_731906;
	.loc	20	410	0
	mov.f32 	%f296, %f12;
	mov.f32 	%f122, %f296;
	mov.f32 	%f297, %f13;
	mov.f32 	%f124, %f297;
	mov.f32 	%f298, %f14;
	mov.f32 	%f126, %f298;
	.loc	20	57	0
	mov.s32 	%r323, 0;
	setp.eq.s32 	%p43, %r41, %r323;
	@!%p43 bra 	$Lt_38_519170;
	.loc	20	59	0
	mov.f32 	%f127, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_101;
$Lt_38_519170:
	.loc	20	61	0
	mov.u32 	%r324, 64;
	setp.ne.s32 	%p217, %r41, %r324;
	@%p217 bra 	$Lt_38_519426;
	.loc	20	63	0
	mov.f32 	%f127, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_101;
$Lt_38_519426:
	.loc	20	65	0
	mov.u32 	%r325, 128;
	setp.ne.s32 	%p218, %r41, %r325;
	@%p218 bra 	$Lt_38_519682;
	.loc	20	68	0
	mov.f32 	%f127, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_101;
$Lt_38_519682:
	.loc	20	70	0
	mov.u32 	%r326, 192;
	setp.ne.s32 	%p219, %r41, %r326;
	@%p219 bra 	$Lt_38_519938;
	.loc	20	72	0
	mov.f32 	%f127, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_101;
$Lt_38_519938:
	.loc	20	76	0
	mov.f32 	%f127, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_101:
	.loc	20	413	0
	mov.u32 	%r327, 0;
	setp.eq.s32 	%p220, %r46, %r327;
	@%p220 bra 	$Lt_38_732162;
	.loc	20	57	0
	@!%p43 bra 	$Lt_38_520450;
	.loc	20	59	0
	mov.f32 	%f128, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_99;
$Lt_38_520450:
	.loc	20	61	0
	mov.u32 	%r328, 64;
	setp.ne.s32 	%p221, %r41, %r328;
	@%p221 bra 	$Lt_38_520706;
	.loc	20	63	0
	mov.f32 	%f128, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_99;
$Lt_38_520706:
	.loc	20	65	0
	mov.u32 	%r329, 128;
	setp.ne.s32 	%p222, %r41, %r329;
	@%p222 bra 	$Lt_38_520962;
	.loc	20	68	0
	mov.f32 	%f128, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_99;
$Lt_38_520962:
	.loc	20	70	0
	mov.u32 	%r330, 192;
	setp.ne.s32 	%p223, %r41, %r330;
	@%p223 bra 	$Lt_38_521218;
	.loc	20	72	0
	mov.f32 	%f128, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_99;
$Lt_38_521218:
	.loc	20	76	0
	mov.f32 	%f128, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_99:
	.loc	20	118	0
	and.b32 	%r148, %r20, 2048;
	mov.s32 	%r331, 0;
	setp.ne.s32 	%p89, %r148, %r331;
	@!%p89 bra 	$Lt_38_732930;
	.loc	20	100	0
	ld.const.f32 	%f299, [kYCbCrOffset+0];
	bra.uni 	$Lt_38_732674;
$Lt_38_732930:
	ld.const.f32 	%f299, [kYCbCrFullRangeOffset+0];
$Lt_38_732674:
	.loc	20	118	0
	@!%p89 bra 	$Lt_38_733442;
	.loc	20	100	0
	ld.const.f32 	%f300, [kYCbCrOffset+4];
	bra.uni 	$Lt_38_733186;
$Lt_38_733442:
	ld.const.f32 	%f300, [kYCbCrFullRangeOffset+4];
$Lt_38_733186:
	.loc	20	118	0
	@!%p89 bra 	$Lt_38_733954;
	.loc	20	100	0
	ld.const.f32 	%f301, [kYCbCrOffset+8];
	bra.uni 	$Lt_38_733698;
$Lt_38_733954:
	ld.const.f32 	%f301, [kYCbCrFullRangeOffset+8];
$Lt_38_733698:
	.loc	20	415	0
	mov.f32 	%f302, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f303, %f128, %f302;
	mul.ftz.f32 	%f304, %f303, %f299;
	sub.ftz.f32 	%f122, %f296, %f304;
	mul.ftz.f32 	%f305, %f303, %f300;
	sub.ftz.f32 	%f124, %f297, %f305;
	mul.ftz.f32 	%f306, %f303, %f301;
	sub.ftz.f32 	%f126, %f298, %f306;
$Lt_38_732162:
	.loc	20	418	0
	rcp.approx.ftz.f32 	%f307, %f127;
	mul.ftz.f32 	%f308, %f307, %f10;
	mul.ftz.f32 	%f122, %f308, %f122;
	.loc	20	419	0
	mul.ftz.f32 	%f124, %f308, %f124;
	.loc	20	420	0
	mul.ftz.f32 	%f126, %f308, %f126;
	.loc	20	422	0
	mov.u32 	%r332, 0;
	setp.eq.s32 	%p224, %r46, %r332;
	@%p224 bra 	$Lt_38_734210;
	.loc	20	57	0
	@!%p43 bra 	$Lt_38_523522;
	.loc	20	59	0
	mov.f32 	%f139, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_95;
$Lt_38_523522:
	.loc	20	61	0
	mov.u32 	%r333, 64;
	setp.ne.s32 	%p225, %r41, %r333;
	@%p225 bra 	$Lt_38_523778;
	.loc	20	63	0
	mov.f32 	%f139, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_95;
$Lt_38_523778:
	.loc	20	65	0
	mov.u32 	%r334, 128;
	setp.ne.s32 	%p226, %r41, %r334;
	@%p226 bra 	$Lt_38_524034;
	.loc	20	68	0
	mov.f32 	%f139, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_95;
$Lt_38_524034:
	.loc	20	70	0
	mov.u32 	%r335, 192;
	setp.ne.s32 	%p227, %r41, %r335;
	@%p227 bra 	$Lt_38_524290;
	.loc	20	72	0
	mov.f32 	%f139, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_95;
$Lt_38_524290:
	.loc	20	76	0
	mov.f32 	%f139, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_95:
	.loc	20	107	0
	and.b32 	%r148, %r20, 2048;
	mov.s32 	%r336, 0;
	setp.ne.s32 	%p89, %r148, %r336;
	@!%p89 bra 	$Lt_38_734978;
	.loc	20	100	0
	ld.const.f32 	%f309, [kYCbCrOffset+0];
	bra.uni 	$Lt_38_734722;
$Lt_38_734978:
	ld.const.f32 	%f309, [kYCbCrFullRangeOffset+0];
$Lt_38_734722:
	.loc	20	107	0
	@!%p89 bra 	$Lt_38_735490;
	.loc	20	100	0
	ld.const.f32 	%f310, [kYCbCrOffset+4];
	bra.uni 	$Lt_38_735234;
$Lt_38_735490:
	ld.const.f32 	%f310, [kYCbCrFullRangeOffset+4];
$Lt_38_735234:
	.loc	20	107	0
	@!%p89 bra 	$Lt_38_736002;
	.loc	20	100	0
	ld.const.f32 	%f311, [kYCbCrOffset+8];
	bra.uni 	$Lt_38_735746;
$Lt_38_736002:
	ld.const.f32 	%f311, [kYCbCrFullRangeOffset+8];
$Lt_38_735746:
	.loc	20	424	0
	mov.f32 	%f312, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f313, %f139, %f312;
	fma.rn.ftz.f32 	%f122, %f313, %f309, %f122;
	fma.rn.ftz.f32 	%f124, %f313, %f310, %f124;
	fma.rn.ftz.f32 	%f126, %f313, %f311, %f126;
$Lt_38_734210:
	.loc	20	543	0
	mov.f32 	%f12, %f122;
	mov.f32 	%f13, %f124;
	mov.f32 	%f14, %f126;
	@!%p99 bra 	$Lt_38_731650;
	.loc	20	57	0
	@!%p43 bra 	$Lt_38_526594;
	.loc	20	59	0
	mov.f32 	%f120, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_91;
$Lt_38_526594:
	.loc	20	61	0
	mov.u32 	%r337, 64;
	setp.ne.s32 	%p228, %r41, %r337;
	@%p228 bra 	$Lt_38_526850;
	.loc	20	63	0
	mov.f32 	%f120, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_91;
$Lt_38_526850:
	.loc	20	65	0
	mov.u32 	%r338, 128;
	setp.ne.s32 	%p229, %r41, %r338;
	@%p229 bra 	$Lt_38_527106;
	.loc	20	68	0
	mov.f32 	%f120, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_91;
$Lt_38_527106:
	.loc	20	70	0
	mov.u32 	%r339, 192;
	setp.ne.s32 	%p230, %r41, %r339;
	@%p230 bra 	$Lt_38_527362;
	.loc	20	72	0
	mov.f32 	%f120, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_91;
$Lt_38_527362:
	.loc	20	76	0
	mov.f32 	%f120, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_91:
	.loc	20	546	0
	mov.f32 	%f10, %f120;
	bra.uni 	$Lt_38_731650;
$Lt_38_731906:
	.loc	20	433	0
	mov.f32 	%f145, %f10;
	mov.f32 	%f314, %f12;
	mov.f32 	%f147, %f314;
	mov.f32 	%f315, %f13;
	mov.f32 	%f149, %f315;
	mov.f32 	%f316, %f14;
	mov.f32 	%f151, %f316;
	.loc	20	435	0
	mov.u32 	%r340, 0;
	setp.eq.s32 	%p231, %r46, %r340;
	@%p231 bra 	$Lt_38_736770;
	.loc	20	57	0
	mov.u32 	%r341, 0;
	setp.ne.s32 	%p232, %r41, %r341;
	@%p232 bra 	$Lt_38_527874;
	.loc	20	59	0
	mov.f32 	%f152, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_89;
$Lt_38_527874:
	.loc	20	61	0
	mov.u32 	%r342, 64;
	setp.ne.s32 	%p233, %r41, %r342;
	@%p233 bra 	$Lt_38_528130;
	.loc	20	63	0
	mov.f32 	%f152, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_89;
$Lt_38_528130:
	.loc	20	65	0
	mov.u32 	%r343, 128;
	setp.ne.s32 	%p234, %r41, %r343;
	@%p234 bra 	$Lt_38_528386;
	.loc	20	68	0
	mov.f32 	%f152, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_89;
$Lt_38_528386:
	.loc	20	70	0
	mov.u32 	%r344, 192;
	setp.ne.s32 	%p235, %r41, %r344;
	@%p235 bra 	$Lt_38_528642;
	.loc	20	72	0
	mov.f32 	%f152, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_89;
$Lt_38_528642:
	.loc	20	76	0
	mov.f32 	%f152, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_89:
	.loc	20	118	0
	and.b32 	%r148, %r20, 2048;
	mov.s32 	%r345, 0;
	setp.ne.s32 	%p89, %r148, %r345;
	@!%p89 bra 	$Lt_38_737538;
	.loc	20	100	0
	ld.const.f32 	%f317, [kYCbCrOffset+0];
	bra.uni 	$Lt_38_737282;
$Lt_38_737538:
	ld.const.f32 	%f317, [kYCbCrFullRangeOffset+0];
$Lt_38_737282:
	.loc	20	118	0
	@!%p89 bra 	$Lt_38_738050;
	.loc	20	100	0
	ld.const.f32 	%f318, [kYCbCrOffset+4];
	bra.uni 	$Lt_38_737794;
$Lt_38_738050:
	ld.const.f32 	%f318, [kYCbCrFullRangeOffset+4];
$Lt_38_737794:
	.loc	20	118	0
	@!%p89 bra 	$Lt_38_738562;
	.loc	20	100	0
	ld.const.f32 	%f319, [kYCbCrOffset+8];
	bra.uni 	$Lt_38_738306;
$Lt_38_738562:
	ld.const.f32 	%f319, [kYCbCrFullRangeOffset+8];
$Lt_38_738306:
	.loc	20	437	0
	mov.f32 	%f320, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f321, %f152, %f320;
	mul.ftz.f32 	%f322, %f321, %f317;
	sub.ftz.f32 	%f147, %f314, %f322;
	mul.ftz.f32 	%f323, %f321, %f318;
	sub.ftz.f32 	%f149, %f315, %f323;
	mul.ftz.f32 	%f324, %f321, %f319;
	sub.ftz.f32 	%f151, %f316, %f324;
$Lt_38_736770:
	mov.f32 	%f325, 0fb70637bd;   	// -8e-006
	add.ftz.f32 	%f326, %f10, %f325;
	mov.f32 	%f327, 0f00000000;   	// 0
	setp.le.ftz.f32 	%p236, %f326, %f327;
	@!%p236 bra 	$Lt_38_739074;
	mov.f32 	%f151, 0f00000000;   	// 0
	mov.f32 	%f149, 0f00000000;   	// 0
	mov.f32 	%f147, 0f00000000;   	// 0
	mov.f32 	%f145, 0f00000000;   	// 0
	bra.uni 	$Lt_38_738818;
$Lt_38_739074:
	.loc	20	57	0
	mov.u32 	%r346, 0;
	setp.ne.s32 	%p237, %r41, %r346;
	@%p237 bra 	$Lt_38_530946;
	.loc	20	59	0
	mov.f32 	%f164, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_85;
$Lt_38_530946:
	.loc	20	61	0
	mov.u32 	%r347, 64;
	setp.ne.s32 	%p238, %r41, %r347;
	@%p238 bra 	$Lt_38_531202;
	.loc	20	63	0
	mov.f32 	%f164, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_85;
$Lt_38_531202:
	.loc	20	65	0
	mov.u32 	%r348, 128;
	setp.ne.s32 	%p239, %r41, %r348;
	@%p239 bra 	$Lt_38_531458;
	.loc	20	68	0
	mov.f32 	%f164, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_85;
$Lt_38_531458:
	.loc	20	70	0
	mov.u32 	%r349, 192;
	setp.ne.s32 	%p240, %r41, %r349;
	@%p240 bra 	$Lt_38_531714;
	.loc	20	72	0
	mov.f32 	%f164, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_85;
$Lt_38_531714:
	.loc	20	76	0
	mov.f32 	%f164, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_85:
	.loc	20	447	0
	div.approx.ftz.f32 	%f328, %f164, %f10;
	mul.ftz.f32 	%f147, %f328, %f147;
	.loc	20	448	0
	mul.ftz.f32 	%f149, %f328, %f149;
	.loc	20	449	0
	mul.ftz.f32 	%f151, %f328, %f151;
$Lt_38_738818:
	.loc	20	452	0
	mov.u32 	%r350, 0;
	setp.eq.s32 	%p241, %r46, %r350;
	@%p241 bra 	$Lt_38_739330;
	.loc	20	57	0
	mov.u32 	%r351, 0;
	setp.ne.s32 	%p242, %r41, %r351;
	@%p242 bra 	$Lt_38_532226;
	.loc	20	59	0
	mov.f32 	%f166, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_83;
$Lt_38_532226:
	.loc	20	61	0
	mov.u32 	%r352, 64;
	setp.ne.s32 	%p243, %r41, %r352;
	@%p243 bra 	$Lt_38_532482;
	.loc	20	63	0
	mov.f32 	%f166, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_83;
$Lt_38_532482:
	.loc	20	65	0
	mov.u32 	%r353, 128;
	setp.ne.s32 	%p244, %r41, %r353;
	@%p244 bra 	$Lt_38_532738;
	.loc	20	68	0
	mov.f32 	%f166, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_83;
$Lt_38_532738:
	.loc	20	70	0
	mov.u32 	%r354, 192;
	setp.ne.s32 	%p245, %r41, %r354;
	@%p245 bra 	$Lt_38_532994;
	.loc	20	72	0
	mov.f32 	%f166, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_83;
$Lt_38_532994:
	.loc	20	76	0
	mov.f32 	%f166, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_83:
	.loc	20	107	0
	and.b32 	%r148, %r20, 2048;
	mov.s32 	%r355, 0;
	setp.ne.s32 	%p89, %r148, %r355;
	@!%p89 bra 	$Lt_38_740098;
	.loc	20	100	0
	ld.const.f32 	%f329, [kYCbCrOffset+0];
	bra.uni 	$Lt_38_739842;
$Lt_38_740098:
	ld.const.f32 	%f329, [kYCbCrFullRangeOffset+0];
$Lt_38_739842:
	.loc	20	107	0
	@!%p89 bra 	$Lt_38_740610;
	.loc	20	100	0
	ld.const.f32 	%f330, [kYCbCrOffset+4];
	bra.uni 	$Lt_38_740354;
$Lt_38_740610:
	ld.const.f32 	%f330, [kYCbCrFullRangeOffset+4];
$Lt_38_740354:
	.loc	20	107	0
	@!%p89 bra 	$Lt_38_741122;
	.loc	20	100	0
	ld.const.f32 	%f331, [kYCbCrOffset+8];
	bra.uni 	$Lt_38_740866;
$Lt_38_741122:
	ld.const.f32 	%f331, [kYCbCrFullRangeOffset+8];
$Lt_38_740866:
	.loc	20	454	0
	mov.f32 	%f332, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f333, %f166, %f332;
	fma.rn.ftz.f32 	%f147, %f333, %f329, %f147;
	fma.rn.ftz.f32 	%f149, %f333, %f330, %f149;
	fma.rn.ftz.f32 	%f151, %f333, %f331, %f151;
$Lt_38_739330:
	.loc	20	551	0
	mov.f32 	%f10, %f145;
	mov.f32 	%f12, %f147;
	mov.f32 	%f13, %f149;
	mov.f32 	%f14, %f151;
$Lt_38_731650:
$L_38_682498:
$Lt_38_731138:
	.loc	20	540	0
	@!%p130 bra 	$Lt_38_741378;
	.loc	21	268	0
	mov.f32 	%f334, %f12;
	.loc	21	269	0
	mov.f32 	%f335, %f10;
	.loc	20	558	0
	mov.f32 	%f10, %f14;
	mov.f32 	%f12, %f13;
	mov.f32 	%f13, %f334;
	mov.f32 	%f14, %f335;
$Lt_38_741378:
	@!%p8 bra 	$Lt_38_741890;
	.loc	20	57	0
	mov.u32 	%r356, 0;
	setp.ne.s32 	%p246, %r41, %r356;
	@%p246 bra 	$Lt_38_535554;
	.loc	20	59	0
	mov.f32 	%f174, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_79;
$Lt_38_535554:
	.loc	20	61	0
	mov.u32 	%r357, 64;
	setp.ne.s32 	%p247, %r41, %r357;
	@%p247 bra 	$Lt_38_535810;
	.loc	20	63	0
	mov.f32 	%f174, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_79;
$Lt_38_535810:
	.loc	20	65	0
	mov.u32 	%r358, 128;
	setp.ne.s32 	%p248, %r41, %r358;
	@%p248 bra 	$Lt_38_536066;
	.loc	20	68	0
	mov.f32 	%f174, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_79;
$Lt_38_536066:
	.loc	20	70	0
	mov.u32 	%r359, 192;
	setp.ne.s32 	%p249, %r41, %r359;
	@%p249 bra 	$Lt_38_536322;
	.loc	20	72	0
	mov.f32 	%f174, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_79;
$Lt_38_536322:
	.loc	20	76	0
	mov.f32 	%f174, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_79:
	.loc	20	564	0
	mov.f32 	%f336, 0f3f000000;   	// 0.5
	add.ftz.f32 	%f337, %f10, %f336;
	mov.f32 	%f338, 0f00000000;   	// 0
	mov.f32 	%f339, 0f00000000;   	// 0
	setp.gt.ftz.f32 	%p250, %f337, %f339;
	selp.f32 	%f340, %f337, %f338, %p250;
	min.ftz.f32 	%f10, %f340, %f174;
	mov.f32 	%f341, 0f3f000000;   	// 0.5
	add.ftz.f32 	%f342, %f12, %f341;
	mov.f32 	%f343, 0f00000000;   	// 0
	mov.f32 	%f344, 0f00000000;   	// 0
	setp.gt.ftz.f32 	%p251, %f342, %f344;
	selp.f32 	%f345, %f342, %f343, %p251;
	min.ftz.f32 	%f12, %f345, %f174;
	mov.f32 	%f346, 0f3f000000;   	// 0.5
	add.ftz.f32 	%f347, %f13, %f346;
	mov.f32 	%f348, 0f00000000;   	// 0
	mov.f32 	%f349, 0f00000000;   	// 0
	setp.gt.ftz.f32 	%p252, %f347, %f349;
	selp.f32 	%f350, %f347, %f348, %p252;
	min.ftz.f32 	%f13, %f350, %f174;
	mov.f32 	%f351, 0f3f000000;   	// 0.5
	add.ftz.f32 	%f352, %f14, %f351;
	mov.f32 	%f353, 0f00000000;   	// 0
	mov.f32 	%f354, 0f00000000;   	// 0
	setp.gt.ftz.f32 	%p253, %f352, %f354;
	selp.f32 	%f355, %f352, %f353, %p253;
	min.ftz.f32 	%f14, %f355, %f174;
$Lt_38_741890:
	st.v4.f32 	[%rd4+0], {%f10,%f12,%f13,%f14};
	.loc	22	67	0
	bra.uni 	$Lt_38_714754;
$Lt_38_715010:
	.loc	20	57	0
	@!%p3 bra 	$Lt_38_538114;
	.loc	20	59	0
	mov.f32 	%f8, 0f437f0000;     	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_77;
$Lt_38_538114:
	.loc	20	61	0
	mov.u32 	%r360, 64;
	setp.ne.s32 	%p254, %r34, %r360;
	@%p254 bra 	$Lt_38_538370;
	.loc	20	63	0
	mov.f32 	%f8, 0f447fc000;     	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_77;
$Lt_38_538370:
	.loc	20	65	0
	mov.u32 	%r361, 128;
	setp.ne.s32 	%p255, %r34, %r361;
	@%p255 bra 	$Lt_38_538626;
	.loc	20	68	0
	mov.f32 	%f8, 0f47000000;     	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_77;
$Lt_38_538626:
	.loc	20	70	0
	mov.u32 	%r362, 192;
	setp.ne.s32 	%p256, %r34, %r362;
	@%p256 bra 	$Lt_38_538882;
	.loc	20	72	0
	mov.f32 	%f8, 0fbf800000;     	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_77;
$Lt_38_538882:
	.loc	20	76	0
	mov.f32 	%f8, 0f3f800000;     	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_77:
	.loc	20	469	0
	cvt.u8.u32 	%r363, %r14;
	cvt.rn.f32.u32 	%f356, %r363;
	add.ftz.f32 	%f357, %f356, %f5;
	cvt.rzi.ftz.u32.f32 	%r364, %f357;
	shr.u32 	%r365, %r364, 1;
	cvt.rn.f32.u32 	%f358, %r365;
	mov.f32 	%f10, %f358;
	cvt.u8.u32 	%r366, %r10;
	cvt.rn.f32.u32 	%f359, %r366;
	add.ftz.f32 	%f360, %f359, %f1;
	cvt.rzi.ftz.u32.f32 	%r367, %f360;
	shr.u32 	%r368, %r367, 1;
	cvt.rn.f32.u32 	%f361, %r368;
	mov.f32 	%f12, %f361;
	mov.f32 	%f13, %f7;
	mov.f32 	%f14, %f8;
	@!%p7 bra 	$Lt_38_742402;
	.loc	20	473	0
	mov.f32 	%f10, %f14;
	mov.f32 	%f12, %f7;
	mov.f32 	%f13, %f361;
	mov.f32 	%f14, %f358;
$Lt_38_742402:
	.loc	20	476	0
	mov.u32 	%r369, 0;
	setp.eq.s32 	%p257, %r51, %r369;
	@%p257 bra 	$Lt_38_779778;
	mov.u32 	%r370, 0;
	setp.ne.s32 	%p258, %r46, %r370;
	@%p258 bra 	$Lt_38_779778;
	.loc	20	57	0
	@!%p3 bra 	$Lt_38_539394;
	.loc	20	59	0
	mov.f32 	%f15, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_75;
$Lt_38_539394:
	.loc	20	61	0
	mov.u32 	%r371, 64;
	setp.ne.s32 	%p259, %r34, %r371;
	@%p259 bra 	$Lt_38_539650;
	.loc	20	63	0
	mov.f32 	%f15, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_75;
$Lt_38_539650:
	.loc	20	65	0
	mov.u32 	%r372, 128;
	setp.ne.s32 	%p260, %r34, %r372;
	@%p260 bra 	$Lt_38_539906;
	.loc	20	68	0
	mov.f32 	%f15, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_75;
$Lt_38_539906:
	.loc	20	70	0
	mov.u32 	%r373, 192;
	setp.ne.s32 	%p261, %r34, %r373;
	@%p261 bra 	$Lt_38_540162;
	.loc	20	72	0
	mov.f32 	%f15, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_75;
$Lt_38_540162:
	.loc	20	76	0
	mov.f32 	%f15, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_75:
	.loc	20	118	0
	and.b32 	%r57, %r33, 2048;
	mov.s32 	%r374, 0;
	setp.ne.s32 	%p15, %r57, %r374;
	@!%p15 bra 	$Lt_38_743170;
	.loc	20	100	0
	ld.const.f32 	%f362, [kYCbCrOffset+0];
	bra.uni 	$Lt_38_742914;
$Lt_38_743170:
	ld.const.f32 	%f362, [kYCbCrFullRangeOffset+0];
$Lt_38_742914:
	.loc	20	118	0
	@!%p15 bra 	$Lt_38_743682;
	.loc	20	100	0
	ld.const.f32 	%f363, [kYCbCrOffset+4];
	bra.uni 	$Lt_38_743426;
$Lt_38_743682:
	ld.const.f32 	%f363, [kYCbCrFullRangeOffset+4];
$Lt_38_743426:
	.loc	20	118	0
	@!%p15 bra 	$Lt_38_744194;
	.loc	20	100	0
	ld.const.f32 	%f364, [kYCbCrOffset+8];
	bra.uni 	$Lt_38_743938;
$Lt_38_744194:
	ld.const.f32 	%f364, [kYCbCrFullRangeOffset+8];
$Lt_38_743938:
	.loc	20	478	0
	mov.f32 	%f365, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f366, %f15, %f365;
	mul.ftz.f32 	%f367, %f366, %f362;
	sub.ftz.f32 	%f12, %f12, %f367;
	mul.ftz.f32 	%f368, %f366, %f363;
	sub.ftz.f32 	%f13, %f13, %f368;
	mul.ftz.f32 	%f369, %f366, %f364;
	sub.ftz.f32 	%f14, %f14, %f369;
$Lt_38_779778:
$Lt_38_317442:
	.loc	20	481	0
	@!%p16 bra 	$Lt_38_780290;
	mov.u32 	%r375, 0;
	setp.ne.s32 	%p262, %r59, %r375;
	@%p262 bra 	$Lt_38_780290;
	.loc	20	483	0
	mov.f32 	%f370, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p263, %f12, %f370;
	@!%p263 bra 	$Lt_38_744450;
	.loc	20	372	0
	neg.ftz.f32 	%f371, %f12;
	lg2.approx.ftz.f32 	%f372, %f371;
	mov.f32 	%f373, 0f3ee66666;   	// 0.45
	mul.ftz.f32 	%f374, %f372, %f373;
	ex2.approx.ftz.f32 	%f375, %f374;
	neg.ftz.f32 	%f30, %f375;
	bra.uni 	$LDWendi___log2f_216_71;
$Lt_38_744450:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f376, %f12;
	mov.f32 	%f377, 0f3ee66666;   	// 0.45
	mul.ftz.f32 	%f378, %f376, %f377;
	ex2.approx.ftz.f32 	%f30, %f378;
$LDWendi___log2f_216_71:
	.loc	20	483	0
	mov.f32 	%f379, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p264, %f13, %f379;
	@!%p264 bra 	$Lt_38_744962;
	.loc	20	372	0
	neg.ftz.f32 	%f380, %f13;
	lg2.approx.ftz.f32 	%f381, %f380;
	mov.f32 	%f382, 0f3ee66666;   	// 0.45
	mul.ftz.f32 	%f383, %f381, %f382;
	ex2.approx.ftz.f32 	%f384, %f383;
	neg.ftz.f32 	%f40, %f384;
	bra.uni 	$LDWendi___log2f_216_69;
$Lt_38_744962:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f385, %f13;
	mov.f32 	%f386, 0f3ee66666;   	// 0.45
	mul.ftz.f32 	%f387, %f385, %f386;
	ex2.approx.ftz.f32 	%f40, %f387;
$LDWendi___log2f_216_69:
	.loc	20	483	0
	mov.f32 	%f388, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p265, %f14, %f388;
	@!%p265 bra 	$Lt_38_745474;
	.loc	20	372	0
	neg.ftz.f32 	%f389, %f14;
	lg2.approx.ftz.f32 	%f390, %f389;
	mov.f32 	%f391, 0f3ee66666;   	// 0.45
	mul.ftz.f32 	%f392, %f390, %f391;
	ex2.approx.ftz.f32 	%f393, %f392;
	neg.ftz.f32 	%f50, %f393;
	bra.uni 	$LDWendi___log2f_216_67;
$Lt_38_745474:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f394, %f14;
	mov.f32 	%f395, 0f3ee66666;   	// 0.45
	mul.ftz.f32 	%f396, %f394, %f395;
	ex2.approx.ftz.f32 	%f50, %f396;
$LDWendi___log2f_216_67:
	.loc	20	483	0
	mov.f32 	%f12, %f30;
	mov.f32 	%f13, %f40;
	mov.f32 	%f14, %f50;
$Lt_38_780290:
$Lt_38_321282:
	.loc	20	486	0
	mov.u32 	%r376, 0;
	setp.ne.s32 	%p266, %r69, %r376;
	@%p266 bra 	$Lt_38_371970;
	setp.eq.s32 	%p267, %r46, %r51;
	@%p267 bra 	$Lt_38_372226;
$Lt_38_371970:
	.loc	20	490	0
	mov.u32 	%r377, 0;
	setp.ne.s32 	%p268, %r65, %r377;
	@%p268 bra 	$Lt_38_746242;
	mov.s32 	%r378, 256;
	setp.eq.s32 	%p26, %r41, %r378;
	mov.u32 	%r379, 256;
	setp.ne.s32 	%p269, %r34, %r379;
	@%p269 bra 	$Lt_38_746754;
	.loc	20	137	0
	mov.s32 	%r380, 512;
	setp.eq.s32 	%p28, %r64, %r380;
	@!%p26 bra 	$Lt_38_543490;
	.loc	20	139	0
	@!%p28 bra 	$Lt_38_543746;
	.loc	20	141	0
	mov.s64 	%rd6, %rd5;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_65;
$Lt_38_543746:
	.loc	20	145	0
	mov.s64 	%rd6, %rd7;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_65;
$Lt_38_543490:
	.loc	20	150	0
	@!%p28 bra 	$Lt_38_544002;
	.loc	20	152	0
	mov.s64 	%rd6, %rd8;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_65;
$Lt_38_544002:
	.loc	20	154	0
	and.b32 	%r381, %r20, 2048;
	mov.u32 	%r382, 0;
	setp.ne.s32 	%p270, %r381, %r382;
	@%p270 bra 	$Lt_38_544258;
	.loc	20	156	0
	mov.s64 	%rd6, %rd9;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_65;
$Lt_38_544258:
	.loc	20	160	0
	mov.s64 	%rd6, %rd10;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_65;
$Lt_38_746754:
	@!%p26 bra 	$Lt_38_747266;
	bra.uni 	$Lt_38_745986;
$Lt_38_747266:
	.loc	20	179	0
	mov.u32 	%r383, 512;
	setp.ne.s32 	%p271, %r64, %r383;
	@%p271 bra 	$Lt_38_545026;
	.loc	20	181	0
	mov.s64 	%rd6, %rd11;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_65;
$Lt_38_545026:
	.loc	20	183	0
	and.b32 	%r384, %r20, 2048;
	mov.u32 	%r385, 0;
	setp.ne.s32 	%p272, %r384, %r385;
	@%p272 bra 	$Lt_38_545282;
	.loc	20	185	0
	mov.s64 	%rd6, %rd12;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_65;
$Lt_38_545282:
	.loc	20	189	0
	mov.s64 	%rd6, %rd13;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_65;
$Lt_38_746242:
	mov.s32 	%r386, 0;
	setp.eq.s32 	%p32, %r63, %r386;
	mov.u32 	%r387, 512;
	setp.ne.s32 	%p273, %r66, %r387;
	@%p273 bra 	$Lt_38_747778;
	mov.s32 	%r388, 256;
	setp.eq.s32 	%p34, %r34, %r388;
	@!%p32 bra 	$Lt_38_748290;
	mov.s32 	%r389, 256;
	setp.eq.s32 	%p26, %r41, %r389;
	@!%p34 bra 	$Lt_38_748802;
	@!%p26 bra 	$Lt_38_745986;
	.loc	20	202	0
	mov.s64 	%rd6, %rd14;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_65;
$Lt_38_748802:
	.loc	20	211	0
	@!%p26 bra 	$Lt_38_546562;
	.loc	20	213	0
	mov.s64 	%rd6, %rd15;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_65;
$Lt_38_546562:
	.loc	20	217	0
	mov.s64 	%rd6, %rd16;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_65;
$Lt_38_748290:
	@!%p34 bra 	$Lt_38_749826;
	bra.uni 	$Lt_38_745986;
$Lt_38_749826:
	mov.s32 	%r390, 256;
	set.eq.u32.s32 	%r391, %r41, %r390;
	neg.s32 	%r392, %r391;
	and.b32 	%r393, %r20, 2048;
	mov.s32 	%r394, 0;
	set.eq.u32.s32 	%r395, %r393, %r394;
	neg.s32 	%r396, %r395;
	or.b32 	%r397, %r392, %r396;
	mov.u32 	%r398, 0;
	setp.eq.s32 	%p274, %r397, %r398;
	@%p274 bra 	$Lt_38_750338;
	bra.uni 	$Lt_38_745986;
$Lt_38_750338:
	.loc	20	250	0
	mov.s64 	%rd6, %rd17;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_65;
$Lt_38_747778:
	and.b32 	%r57, %r33, 2048;
	mov.s32 	%r399, 0;
	setp.eq.s32 	%p275, %r57, %r399;
	mov.pred 	%p37, %p275;
	mov.pred 	%p276, %p39;
	@!%p275 bra 	$Lt_38_750850;
	@!%p32 bra 	$Lt_38_745986;
	.loc	20	259	0
	mov.u32 	%r400, 256;
	setp.ne.s32 	%p277, %r41, %r400;
	@%p277 bra 	$Lt_38_548098;
	.loc	20	261	0
	mov.s64 	%rd6, %rd18;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_65;
$Lt_38_548098:
	.loc	20	265	0
	mov.s64 	%rd6, %rd19;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_65;
$Lt_38_750850:
	mov.s32 	%r401, 256;
	setp.eq.s32 	%p34, %r34, %r401;
	@!%p32 bra 	$Lt_38_751874;
	mov.s32 	%r402, 256;
	setp.eq.s32 	%p26, %r41, %r402;
	@!%p34 bra 	$Lt_38_752386;
	@!%p26 bra 	$Lt_38_745986;
	.loc	20	302	0
	mov.s64 	%rd6, %rd20;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_65;
$Lt_38_752386:
	.loc	20	311	0
	@!%p26 bra 	$Lt_38_549890;
	.loc	20	313	0
	mov.s64 	%rd6, %rd21;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_65;
$Lt_38_549890:
	.loc	20	317	0
	mov.s64 	%rd6, %rd22;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_65;
$Lt_38_751874:
	@!%p34 bra 	$Lt_38_753410;
	bra.uni 	$Lt_38_745986;
$Lt_38_753410:
	selp.s32 	%r403, 1, 0, %p37;
	mov.s32 	%r404, 256;
	set.eq.u32.s32 	%r405, %r41, %r404;
	neg.s32 	%r406, %r405;
	or.b32 	%r407, %r403, %r406;
	mov.u32 	%r408, 0;
	setp.eq.s32 	%p278, %r407, %r408;
	@%p278 bra 	$Lt_38_753922;
	bra.uni 	$Lt_38_745986;
$Lt_38_753922:
	.loc	20	350	0
	mov.s64 	%rd6, %rd23;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_65;
$Lt_38_745986:
	.loc	20	355	0
	mov.u64 	%rd6, 0;
$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__216_65:
	.loc	20	490	0
	ld.f32 	%f397, [%rd6+16];
	mul.ftz.f32 	%f398, %f397, %f13;
	ld.f32 	%f399, [%rd6+12];
	fma.rn.ftz.f32 	%f400, %f399, %f12, %f398;
	ld.f32 	%f401, [%rd6+20];
	fma.rn.ftz.f32 	%f402, %f401, %f14, %f400;
	ld.f32 	%f403, [%rd6+28];
	mul.ftz.f32 	%f404, %f403, %f13;
	ld.f32 	%f405, [%rd6+24];
	fma.rn.ftz.f32 	%f406, %f405, %f12, %f404;
	ld.f32 	%f407, [%rd6+32];
	fma.rn.ftz.f32 	%f408, %f407, %f14, %f406;
	ld.f32 	%f409, [%rd6+4];
	mul.ftz.f32 	%f410, %f409, %f13;
	ld.f32 	%f411, [%rd6+0];
	fma.rn.ftz.f32 	%f412, %f411, %f12, %f410;
	ld.f32 	%f413, [%rd6+8];
	fma.rn.ftz.f32 	%f12, %f413, %f14, %f412;
	mov.f32 	%f13, %f402;
	mov.f32 	%f14, %f408;
	@!%p42 bra 	$Lt_38_754690;
	.loc	20	494	0
	mov.s32 	%r409, 256;
	setp.eq.s32 	%p34, %r34, %r409;
	@!%p34 bra 	$L_38_684546;
	mov.s32 	%r410, 0;
	setp.eq.s32 	%p43, %r41, %r410;
	@%p43 bra 	$L_38_684290;
$L_38_684546:
	@!%p3 bra 	$Lt_38_781570;
	mov.u32 	%r411, 256;
	setp.ne.s32 	%p279, %r41, %r411;
	@%p279 bra 	$Lt_38_781570;
	mov.s32 	%r412, 0;
	setp.eq.s32 	%p43, %r41, %r412;
$L_38_684290:
	.loc	20	57	0
	@!%p43 bra 	$Lt_38_551426;
	.loc	20	59	0
	mov.f32 	%f71, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_63;
$Lt_38_551426:
	.loc	20	61	0
	mov.u32 	%r413, 64;
	setp.ne.s32 	%p280, %r41, %r413;
	@%p280 bra 	$Lt_38_551682;
	.loc	20	63	0
	mov.f32 	%f71, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_63;
$Lt_38_551682:
	.loc	20	65	0
	mov.u32 	%r414, 128;
	setp.ne.s32 	%p281, %r41, %r414;
	@%p281 bra 	$Lt_38_551938;
	.loc	20	68	0
	mov.f32 	%f71, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_63;
$Lt_38_551938:
	.loc	20	70	0
	mov.u32 	%r415, 192;
	setp.ne.s32 	%p282, %r41, %r415;
	@%p282 bra 	$Lt_38_552194;
	.loc	20	72	0
	mov.f32 	%f71, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_63;
$Lt_38_552194:
	.loc	20	76	0
	mov.f32 	%f71, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_63:
	.loc	20	57	0
	@!%p3 bra 	$Lt_38_552450;
	.loc	20	59	0
	mov.f32 	%f72, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_61;
$Lt_38_552450:
	.loc	20	61	0
	mov.u32 	%r416, 64;
	setp.ne.s32 	%p283, %r34, %r416;
	@%p283 bra 	$Lt_38_552706;
	.loc	20	63	0
	mov.f32 	%f72, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_61;
$Lt_38_552706:
	.loc	20	65	0
	mov.u32 	%r417, 128;
	setp.ne.s32 	%p284, %r34, %r417;
	@%p284 bra 	$Lt_38_552962;
	.loc	20	68	0
	mov.f32 	%f72, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_61;
$Lt_38_552962:
	.loc	20	70	0
	mov.u32 	%r418, 192;
	setp.ne.s32 	%p285, %r34, %r418;
	@%p285 bra 	$Lt_38_553218;
	.loc	20	72	0
	mov.f32 	%f72, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_61;
$Lt_38_553218:
	.loc	20	76	0
	mov.f32 	%f72, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_61:
	.loc	20	498	0
	div.approx.ftz.f32 	%f414, %f71, %f72;
	mul.ftz.f32 	%f10, %f10, %f414;
	bra.uni 	$Lt_38_754690;
$Lt_38_781570:
$L_38_684034:
	.loc	20	500	0
	@!%p34 bra 	$L_38_685570;
	@%p8 bra 	$L_38_685314;
$L_38_685570:
	@!%p9 bra 	$Lt_38_782594;
	mov.u32 	%r419, 256;
	setp.ne.s32 	%p286, %r41, %r419;
	@%p286 bra 	$Lt_38_782594;
$L_38_685314:
	.loc	20	57	0
	mov.u32 	%r420, 0;
	setp.ne.s32 	%p287, %r41, %r420;
	@%p287 bra 	$Lt_38_553730;
	.loc	20	59	0
	mov.f32 	%f71, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_59;
$Lt_38_553730:
	.loc	20	61	0
	mov.u32 	%r421, 64;
	setp.ne.s32 	%p288, %r41, %r421;
	@%p288 bra 	$Lt_38_553986;
	.loc	20	63	0
	mov.f32 	%f71, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_59;
$Lt_38_553986:
	.loc	20	65	0
	mov.u32 	%r422, 128;
	setp.ne.s32 	%p289, %r41, %r422;
	@%p289 bra 	$Lt_38_554242;
	.loc	20	68	0
	mov.f32 	%f71, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_59;
$Lt_38_554242:
	.loc	20	70	0
	mov.u32 	%r423, 192;
	setp.ne.s32 	%p290, %r41, %r423;
	@%p290 bra 	$Lt_38_554498;
	.loc	20	72	0
	mov.f32 	%f71, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_59;
$Lt_38_554498:
	.loc	20	76	0
	mov.f32 	%f71, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_59:
	.loc	20	57	0
	@!%p3 bra 	$Lt_38_554754;
	.loc	20	59	0
	mov.f32 	%f72, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_57;
$Lt_38_554754:
	.loc	20	61	0
	mov.u32 	%r424, 64;
	setp.ne.s32 	%p291, %r34, %r424;
	@%p291 bra 	$Lt_38_555010;
	.loc	20	63	0
	mov.f32 	%f72, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_57;
$Lt_38_555010:
	.loc	20	65	0
	mov.u32 	%r425, 128;
	setp.ne.s32 	%p292, %r34, %r425;
	@%p292 bra 	$Lt_38_555266;
	.loc	20	68	0
	mov.f32 	%f72, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_57;
$Lt_38_555266:
	.loc	20	70	0
	mov.u32 	%r426, 192;
	setp.ne.s32 	%p293, %r34, %r426;
	@%p293 bra 	$Lt_38_555522;
	.loc	20	72	0
	mov.f32 	%f72, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_57;
$Lt_38_555522:
	.loc	20	76	0
	mov.f32 	%f72, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_57:
	.loc	20	504	0
	div.approx.ftz.f32 	%f415, %f71, %f72;
	mul.ftz.f32 	%f10, %f10, %f415;
	.loc	20	57	0
	@!%p3 bra 	$Lt_38_556802;
	.loc	20	59	0
	mov.f32 	%f72, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_53;
$Lt_38_556802:
	.loc	20	61	0
	mov.u32 	%r427, 64;
	setp.ne.s32 	%p294, %r34, %r427;
	@%p294 bra 	$Lt_38_557058;
	.loc	20	63	0
	mov.f32 	%f72, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_53;
$Lt_38_557058:
	.loc	20	65	0
	mov.u32 	%r428, 128;
	setp.ne.s32 	%p295, %r34, %r428;
	@%p295 bra 	$Lt_38_557314;
	.loc	20	68	0
	mov.f32 	%f72, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_53;
$Lt_38_557314:
	.loc	20	70	0
	mov.u32 	%r429, 192;
	setp.ne.s32 	%p296, %r34, %r429;
	@%p296 bra 	$Lt_38_557570;
	.loc	20	72	0
	mov.f32 	%f72, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_53;
$Lt_38_557570:
	.loc	20	76	0
	mov.f32 	%f72, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_53:
	.loc	20	505	0
	mov.f32 	%f416, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f417, %f416, %f72;
	mul.ftz.f32 	%f12, %f12, %f417;
	.loc	20	57	0
	@!%p3 bra 	$Lt_38_558850;
	.loc	20	59	0
	mov.f32 	%f72, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_49;
$Lt_38_558850:
	.loc	20	61	0
	mov.u32 	%r430, 64;
	setp.ne.s32 	%p297, %r34, %r430;
	@%p297 bra 	$Lt_38_559106;
	.loc	20	63	0
	mov.f32 	%f72, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_49;
$Lt_38_559106:
	.loc	20	65	0
	mov.u32 	%r431, 128;
	setp.ne.s32 	%p298, %r34, %r431;
	@%p298 bra 	$Lt_38_559362;
	.loc	20	68	0
	mov.f32 	%f72, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_49;
$Lt_38_559362:
	.loc	20	70	0
	mov.u32 	%r432, 192;
	setp.ne.s32 	%p299, %r34, %r432;
	@%p299 bra 	$Lt_38_559618;
	.loc	20	72	0
	mov.f32 	%f72, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_49;
$Lt_38_559618:
	.loc	20	76	0
	mov.f32 	%f72, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_49:
	.loc	20	506	0
	mov.f32 	%f418, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f419, %f418, %f72;
	mul.ftz.f32 	%f13, %f402, %f419;
	.loc	20	57	0
	@!%p3 bra 	$Lt_38_560898;
	.loc	20	59	0
	mov.f32 	%f72, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_45;
$Lt_38_560898:
	.loc	20	61	0
	mov.u32 	%r433, 64;
	setp.ne.s32 	%p300, %r34, %r433;
	@%p300 bra 	$Lt_38_561154;
	.loc	20	63	0
	mov.f32 	%f72, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_45;
$Lt_38_561154:
	.loc	20	65	0
	mov.u32 	%r434, 128;
	setp.ne.s32 	%p301, %r34, %r434;
	@%p301 bra 	$Lt_38_561410;
	.loc	20	68	0
	mov.f32 	%f72, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_45;
$Lt_38_561410:
	.loc	20	70	0
	mov.u32 	%r435, 192;
	setp.ne.s32 	%p302, %r34, %r435;
	@%p302 bra 	$Lt_38_561666;
	.loc	20	72	0
	mov.f32 	%f72, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_45;
$Lt_38_561666:
	.loc	20	76	0
	mov.f32 	%f72, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_45:
	.loc	20	507	0
	mov.f32 	%f420, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f421, %f420, %f72;
	mul.ftz.f32 	%f14, %f408, %f421;
	bra.uni 	$Lt_38_754690;
$Lt_38_782594:
$L_38_685058:
	.loc	20	57	0
	@!%p3 bra 	$Lt_38_561922;
	.loc	20	59	0
	mov.f32 	%f71, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_43;
$Lt_38_561922:
	.loc	20	61	0
	mov.u32 	%r436, 64;
	setp.ne.s32 	%p303, %r34, %r436;
	@%p303 bra 	$Lt_38_562178;
	.loc	20	63	0
	mov.f32 	%f71, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_43;
$Lt_38_562178:
	.loc	20	65	0
	mov.u32 	%r437, 128;
	setp.ne.s32 	%p304, %r34, %r437;
	@%p304 bra 	$Lt_38_562434;
	.loc	20	68	0
	mov.f32 	%f71, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_43;
$Lt_38_562434:
	.loc	20	70	0
	mov.u32 	%r438, 192;
	setp.ne.s32 	%p305, %r34, %r438;
	@%p305 bra 	$Lt_38_562690;
	.loc	20	72	0
	mov.f32 	%f71, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_43;
$Lt_38_562690:
	.loc	20	76	0
	mov.f32 	%f71, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_43:
	.loc	20	511	0
	mov.f32 	%f422, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f423, %f71, %f422;
	mul.ftz.f32 	%f10, %f423, %f10;
	mul.ftz.f32 	%f12, %f423, %f12;
	mul.ftz.f32 	%f13, %f423, %f402;
	mul.ftz.f32 	%f14, %f423, %f408;
	bra.uni 	$Lt_38_754690;
$Lt_38_372226:
	.loc	20	486	0
	@!%p42 bra 	$Lt_38_754690;
	.loc	20	57	0
	mov.u32 	%r439, 0;
	setp.ne.s32 	%p306, %r41, %r439;
	@%p306 bra 	$Lt_38_564226;
	.loc	20	59	0
	mov.f32 	%f71, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_39;
$Lt_38_564226:
	.loc	20	61	0
	mov.u32 	%r440, 64;
	setp.ne.s32 	%p307, %r41, %r440;
	@%p307 bra 	$Lt_38_564482;
	.loc	20	63	0
	mov.f32 	%f71, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_39;
$Lt_38_564482:
	.loc	20	65	0
	mov.u32 	%r441, 128;
	setp.ne.s32 	%p308, %r41, %r441;
	@%p308 bra 	$Lt_38_564738;
	.loc	20	68	0
	mov.f32 	%f71, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_39;
$Lt_38_564738:
	.loc	20	70	0
	mov.u32 	%r442, 192;
	setp.ne.s32 	%p309, %r41, %r442;
	@%p309 bra 	$Lt_38_564994;
	.loc	20	72	0
	mov.f32 	%f71, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_39;
$Lt_38_564994:
	.loc	20	76	0
	mov.f32 	%f71, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_39:
	.loc	20	57	0
	@!%p3 bra 	$Lt_38_565250;
	.loc	20	59	0
	mov.f32 	%f72, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_37;
$Lt_38_565250:
	.loc	20	61	0
	mov.u32 	%r443, 64;
	setp.ne.s32 	%p310, %r34, %r443;
	@%p310 bra 	$Lt_38_565506;
	.loc	20	63	0
	mov.f32 	%f72, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_37;
$Lt_38_565506:
	.loc	20	65	0
	mov.u32 	%r444, 128;
	setp.ne.s32 	%p311, %r34, %r444;
	@%p311 bra 	$Lt_38_565762;
	.loc	20	68	0
	mov.f32 	%f72, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_37;
$Lt_38_565762:
	.loc	20	70	0
	mov.u32 	%r445, 192;
	setp.ne.s32 	%p312, %r34, %r445;
	@%p312 bra 	$Lt_38_566018;
	.loc	20	72	0
	mov.f32 	%f72, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_37;
$Lt_38_566018:
	.loc	20	76	0
	mov.f32 	%f72, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_37:
	.loc	20	517	0
	div.approx.ftz.f32 	%f424, %f71, %f72;
	mul.ftz.f32 	%f10, %f424, %f10;
	mul.ftz.f32 	%f12, %f424, %f12;
	mul.ftz.f32 	%f13, %f424, %f13;
	mul.ftz.f32 	%f14, %f424, %f14;
$Lt_38_754690:
$Lt_38_374530:
	.loc	20	520	0
	@!%p78 bra 	$Lt_38_783106;
	mov.u32 	%r446, 0;
	setp.ne.s32 	%p313, %r60, %r446;
	@%p313 bra 	$Lt_38_783106;
	.loc	20	522	0
	mov.f32 	%f425, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p314, %f12, %f425;
	@!%p314 bra 	$Lt_38_755202;
	.loc	20	372	0
	neg.ftz.f32 	%f426, %f12;
	lg2.approx.ftz.f32 	%f427, %f426;
	mov.f32 	%f428, 0f400e38e4;   	// 2.22222
	mul.ftz.f32 	%f429, %f427, %f428;
	ex2.approx.ftz.f32 	%f430, %f429;
	neg.ftz.f32 	%f90, %f430;
	bra.uni 	$LDWendi___log2f_216_35;
$Lt_38_755202:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f431, %f12;
	mov.f32 	%f432, 0f400e38e4;   	// 2.22222
	mul.ftz.f32 	%f433, %f431, %f432;
	ex2.approx.ftz.f32 	%f90, %f433;
$LDWendi___log2f_216_35:
	.loc	20	522	0
	mov.f32 	%f434, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p315, %f13, %f434;
	@!%p315 bra 	$Lt_38_755714;
	.loc	20	372	0
	neg.ftz.f32 	%f435, %f13;
	lg2.approx.ftz.f32 	%f436, %f435;
	mov.f32 	%f437, 0f400e38e4;   	// 2.22222
	mul.ftz.f32 	%f438, %f436, %f437;
	ex2.approx.ftz.f32 	%f439, %f438;
	neg.ftz.f32 	%f100, %f439;
	bra.uni 	$LDWendi___log2f_216_33;
$Lt_38_755714:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f440, %f13;
	mov.f32 	%f441, 0f400e38e4;   	// 2.22222
	mul.ftz.f32 	%f442, %f440, %f441;
	ex2.approx.ftz.f32 	%f100, %f442;
$LDWendi___log2f_216_33:
	.loc	20	522	0
	mov.f32 	%f443, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p316, %f14, %f443;
	@!%p316 bra 	$Lt_38_756226;
	.loc	20	372	0
	neg.ftz.f32 	%f444, %f14;
	lg2.approx.ftz.f32 	%f445, %f444;
	mov.f32 	%f446, 0f400e38e4;   	// 2.22222
	mul.ftz.f32 	%f447, %f445, %f446;
	ex2.approx.ftz.f32 	%f448, %f447;
	neg.ftz.f32 	%f110, %f448;
	bra.uni 	$LDWendi___log2f_216_31;
$Lt_38_756226:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f449, %f14;
	mov.f32 	%f450, 0f400e38e4;   	// 2.22222
	mul.ftz.f32 	%f451, %f449, %f450;
	ex2.approx.ftz.f32 	%f110, %f451;
$LDWendi___log2f_216_31:
	.loc	20	522	0
	mov.f32 	%f12, %f90;
	mov.f32 	%f13, %f100;
	mov.f32 	%f14, %f110;
$Lt_38_783106:
$Lt_38_376578:
	.loc	20	525	0
	mov.u32 	%r447, 0;
	setp.eq.s32 	%p317, %r46, %r447;
	@%p317 bra 	$Lt_38_783618;
	mov.u32 	%r448, 0;
	setp.ne.s32 	%p318, %r51, %r448;
	@%p318 bra 	$Lt_38_783618;
	.loc	20	57	0
	mov.u32 	%r449, 0;
	setp.ne.s32 	%p319, %r41, %r449;
	@%p319 bra 	$Lt_38_567042;
	.loc	20	59	0
	mov.f32 	%f114, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_29;
$Lt_38_567042:
	.loc	20	61	0
	mov.u32 	%r450, 64;
	setp.ne.s32 	%p320, %r41, %r450;
	@%p320 bra 	$Lt_38_567298;
	.loc	20	63	0
	mov.f32 	%f114, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_29;
$Lt_38_567298:
	.loc	20	65	0
	mov.u32 	%r451, 128;
	setp.ne.s32 	%p321, %r41, %r451;
	@%p321 bra 	$Lt_38_567554;
	.loc	20	68	0
	mov.f32 	%f114, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_29;
$Lt_38_567554:
	.loc	20	70	0
	mov.u32 	%r452, 192;
	setp.ne.s32 	%p322, %r41, %r452;
	@%p322 bra 	$Lt_38_567810;
	.loc	20	72	0
	mov.f32 	%f114, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_29;
$Lt_38_567810:
	.loc	20	76	0
	mov.f32 	%f114, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_29:
	.loc	20	107	0
	and.b32 	%r148, %r20, 2048;
	mov.s32 	%r453, 0;
	setp.ne.s32 	%p89, %r148, %r453;
	@!%p89 bra 	$Lt_38_756994;
	.loc	20	100	0
	ld.const.f32 	%f452, [kYCbCrOffset+0];
	bra.uni 	$Lt_38_756738;
$Lt_38_756994:
	ld.const.f32 	%f452, [kYCbCrFullRangeOffset+0];
$Lt_38_756738:
	.loc	20	107	0
	@!%p89 bra 	$Lt_38_757506;
	.loc	20	100	0
	ld.const.f32 	%f453, [kYCbCrOffset+4];
	bra.uni 	$Lt_38_757250;
$Lt_38_757506:
	ld.const.f32 	%f453, [kYCbCrFullRangeOffset+4];
$Lt_38_757250:
	.loc	20	107	0
	@!%p89 bra 	$Lt_38_758018;
	.loc	20	100	0
	ld.const.f32 	%f454, [kYCbCrOffset+8];
	bra.uni 	$Lt_38_757762;
$Lt_38_758018:
	ld.const.f32 	%f454, [kYCbCrFullRangeOffset+8];
$Lt_38_757762:
	.loc	20	527	0
	mov.f32 	%f455, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f456, %f114, %f455;
	fma.rn.ftz.f32 	%f12, %f456, %f452, %f12;
	fma.rn.ftz.f32 	%f13, %f456, %f453, %f13;
	fma.rn.ftz.f32 	%f14, %f456, %f454, %f14;
$Lt_38_783618:
$Lt_38_382978:
	.loc	20	525	0
	@!%p90 bra 	$Lt_38_758786;
	.loc	20	532	0
	mov.u32 	%r454, 8;
	setp.ne.s32 	%p323, %r151, %r454;
	@%p323 bra 	$L_38_686850;
	mov.u32 	%r455, 12;
	setp.eq.s32 	%p324, %r150, %r455;
	@%p324 bra 	$Lt_38_784386;
$L_38_686850:
	mov.u32 	%r456, 12;
	setp.eq.s32 	%p325, %r151, %r456;
	@%p325 bra 	$Lt_38_784386;
	mov.u32 	%r457, 0;
	setp.ne.s32 	%p326, %r151, %r457;
	@%p326 bra 	$L_38_686082;
$Lt_38_784386:
$L_38_686338:
	.loc	20	57	0
	mov.u32 	%r458, 0;
	setp.ne.s32 	%p327, %r41, %r458;
	@%p327 bra 	$Lt_38_570370;
	.loc	20	59	0
	mov.f32 	%f120, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_25;
$Lt_38_570370:
	.loc	20	61	0
	mov.u32 	%r459, 64;
	setp.ne.s32 	%p328, %r41, %r459;
	@%p328 bra 	$Lt_38_570626;
	.loc	20	63	0
	mov.f32 	%f120, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_25;
$Lt_38_570626:
	.loc	20	65	0
	mov.u32 	%r460, 128;
	setp.ne.s32 	%p329, %r41, %r460;
	@%p329 bra 	$Lt_38_570882;
	.loc	20	68	0
	mov.f32 	%f120, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_25;
$Lt_38_570882:
	.loc	20	70	0
	mov.u32 	%r461, 192;
	setp.ne.s32 	%p330, %r41, %r461;
	@%p330 bra 	$Lt_38_571138;
	.loc	20	72	0
	mov.f32 	%f120, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_25;
$Lt_38_571138:
	.loc	20	76	0
	mov.f32 	%f120, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_25:
	.loc	20	536	0
	mov.f32 	%f10, %f120;
	bra.uni 	$Lt_38_758786;
$L_38_686082:
	.loc	20	540	0
	mov.s32 	%r462, 12;
	setp.eq.s32 	%p99, %r150, %r462;
	mov.s32 	%r463, 4;
	set.eq.u32.s32 	%r464, %r151, %r463;
	neg.s32 	%r465, %r464;
	selp.s32 	%r466, 1, 0, %p99;
	mov.s32 	%r467, 8;
	set.eq.u32.s32 	%r468, %r150, %r467;
	neg.s32 	%r469, %r468;
	or.b32 	%r470, %r466, %r469;
	and.b32 	%r471, %r465, %r470;
	mov.u32 	%r472, 0;
	setp.eq.s32 	%p331, %r471, %r472;
	@%p331 bra 	$Lt_38_759042;
	.loc	20	410	0
	mov.f32 	%f457, %f12;
	mov.f32 	%f122, %f457;
	mov.f32 	%f458, %f13;
	mov.f32 	%f124, %f458;
	mov.f32 	%f459, %f14;
	mov.f32 	%f126, %f459;
	.loc	20	57	0
	mov.s32 	%r473, 0;
	setp.eq.s32 	%p43, %r41, %r473;
	@!%p43 bra 	$Lt_38_571650;
	.loc	20	59	0
	mov.f32 	%f127, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_23;
$Lt_38_571650:
	.loc	20	61	0
	mov.u32 	%r474, 64;
	setp.ne.s32 	%p332, %r41, %r474;
	@%p332 bra 	$Lt_38_571906;
	.loc	20	63	0
	mov.f32 	%f127, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_23;
$Lt_38_571906:
	.loc	20	65	0
	mov.u32 	%r475, 128;
	setp.ne.s32 	%p333, %r41, %r475;
	@%p333 bra 	$Lt_38_572162;
	.loc	20	68	0
	mov.f32 	%f127, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_23;
$Lt_38_572162:
	.loc	20	70	0
	mov.u32 	%r476, 192;
	setp.ne.s32 	%p334, %r41, %r476;
	@%p334 bra 	$Lt_38_572418;
	.loc	20	72	0
	mov.f32 	%f127, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_23;
$Lt_38_572418:
	.loc	20	76	0
	mov.f32 	%f127, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_23:
	.loc	20	413	0
	mov.u32 	%r477, 0;
	setp.eq.s32 	%p335, %r46, %r477;
	@%p335 bra 	$Lt_38_759298;
	.loc	20	57	0
	@!%p43 bra 	$Lt_38_572930;
	.loc	20	59	0
	mov.f32 	%f128, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_21;
$Lt_38_572930:
	.loc	20	61	0
	mov.u32 	%r478, 64;
	setp.ne.s32 	%p336, %r41, %r478;
	@%p336 bra 	$Lt_38_573186;
	.loc	20	63	0
	mov.f32 	%f128, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_21;
$Lt_38_573186:
	.loc	20	65	0
	mov.u32 	%r479, 128;
	setp.ne.s32 	%p337, %r41, %r479;
	@%p337 bra 	$Lt_38_573442;
	.loc	20	68	0
	mov.f32 	%f128, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_21;
$Lt_38_573442:
	.loc	20	70	0
	mov.u32 	%r480, 192;
	setp.ne.s32 	%p338, %r41, %r480;
	@%p338 bra 	$Lt_38_573698;
	.loc	20	72	0
	mov.f32 	%f128, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_21;
$Lt_38_573698:
	.loc	20	76	0
	mov.f32 	%f128, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_21:
	.loc	20	118	0
	and.b32 	%r148, %r20, 2048;
	mov.s32 	%r481, 0;
	setp.ne.s32 	%p89, %r148, %r481;
	@!%p89 bra 	$Lt_38_760066;
	.loc	20	100	0
	ld.const.f32 	%f460, [kYCbCrOffset+0];
	bra.uni 	$Lt_38_759810;
$Lt_38_760066:
	ld.const.f32 	%f460, [kYCbCrFullRangeOffset+0];
$Lt_38_759810:
	.loc	20	118	0
	@!%p89 bra 	$Lt_38_760578;
	.loc	20	100	0
	ld.const.f32 	%f461, [kYCbCrOffset+4];
	bra.uni 	$Lt_38_760322;
$Lt_38_760578:
	ld.const.f32 	%f461, [kYCbCrFullRangeOffset+4];
$Lt_38_760322:
	.loc	20	118	0
	@!%p89 bra 	$Lt_38_761090;
	.loc	20	100	0
	ld.const.f32 	%f462, [kYCbCrOffset+8];
	bra.uni 	$Lt_38_760834;
$Lt_38_761090:
	ld.const.f32 	%f462, [kYCbCrFullRangeOffset+8];
$Lt_38_760834:
	.loc	20	415	0
	mov.f32 	%f463, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f464, %f128, %f463;
	mul.ftz.f32 	%f465, %f464, %f460;
	sub.ftz.f32 	%f122, %f457, %f465;
	mul.ftz.f32 	%f466, %f464, %f461;
	sub.ftz.f32 	%f124, %f458, %f466;
	mul.ftz.f32 	%f467, %f464, %f462;
	sub.ftz.f32 	%f126, %f459, %f467;
$Lt_38_759298:
	.loc	20	418	0
	rcp.approx.ftz.f32 	%f468, %f127;
	mul.ftz.f32 	%f469, %f468, %f10;
	mul.ftz.f32 	%f122, %f469, %f122;
	.loc	20	419	0
	mul.ftz.f32 	%f124, %f469, %f124;
	.loc	20	420	0
	mul.ftz.f32 	%f126, %f469, %f126;
	.loc	20	422	0
	mov.u32 	%r482, 0;
	setp.eq.s32 	%p339, %r46, %r482;
	@%p339 bra 	$Lt_38_761346;
	.loc	20	57	0
	@!%p43 bra 	$Lt_38_576002;
	.loc	20	59	0
	mov.f32 	%f139, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_17;
$Lt_38_576002:
	.loc	20	61	0
	mov.u32 	%r483, 64;
	setp.ne.s32 	%p340, %r41, %r483;
	@%p340 bra 	$Lt_38_576258;
	.loc	20	63	0
	mov.f32 	%f139, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_17;
$Lt_38_576258:
	.loc	20	65	0
	mov.u32 	%r484, 128;
	setp.ne.s32 	%p341, %r41, %r484;
	@%p341 bra 	$Lt_38_576514;
	.loc	20	68	0
	mov.f32 	%f139, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_17;
$Lt_38_576514:
	.loc	20	70	0
	mov.u32 	%r485, 192;
	setp.ne.s32 	%p342, %r41, %r485;
	@%p342 bra 	$Lt_38_576770;
	.loc	20	72	0
	mov.f32 	%f139, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_17;
$Lt_38_576770:
	.loc	20	76	0
	mov.f32 	%f139, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_17:
	.loc	20	107	0
	and.b32 	%r148, %r20, 2048;
	mov.s32 	%r486, 0;
	setp.ne.s32 	%p89, %r148, %r486;
	@!%p89 bra 	$Lt_38_762114;
	.loc	20	100	0
	ld.const.f32 	%f470, [kYCbCrOffset+0];
	bra.uni 	$Lt_38_761858;
$Lt_38_762114:
	ld.const.f32 	%f470, [kYCbCrFullRangeOffset+0];
$Lt_38_761858:
	.loc	20	107	0
	@!%p89 bra 	$Lt_38_762626;
	.loc	20	100	0
	ld.const.f32 	%f471, [kYCbCrOffset+4];
	bra.uni 	$Lt_38_762370;
$Lt_38_762626:
	ld.const.f32 	%f471, [kYCbCrFullRangeOffset+4];
$Lt_38_762370:
	.loc	20	107	0
	@!%p89 bra 	$Lt_38_763138;
	.loc	20	100	0
	ld.const.f32 	%f472, [kYCbCrOffset+8];
	bra.uni 	$Lt_38_762882;
$Lt_38_763138:
	ld.const.f32 	%f472, [kYCbCrFullRangeOffset+8];
$Lt_38_762882:
	.loc	20	424	0
	mov.f32 	%f473, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f474, %f139, %f473;
	fma.rn.ftz.f32 	%f122, %f474, %f470, %f122;
	fma.rn.ftz.f32 	%f124, %f474, %f471, %f124;
	fma.rn.ftz.f32 	%f126, %f474, %f472, %f126;
$Lt_38_761346:
	.loc	20	543	0
	mov.f32 	%f12, %f122;
	mov.f32 	%f13, %f124;
	mov.f32 	%f14, %f126;
	@!%p99 bra 	$Lt_38_758786;
	.loc	20	57	0
	@!%p43 bra 	$Lt_38_579074;
	.loc	20	59	0
	mov.f32 	%f120, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_13;
$Lt_38_579074:
	.loc	20	61	0
	mov.u32 	%r487, 64;
	setp.ne.s32 	%p343, %r41, %r487;
	@%p343 bra 	$Lt_38_579330;
	.loc	20	63	0
	mov.f32 	%f120, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_13;
$Lt_38_579330:
	.loc	20	65	0
	mov.u32 	%r488, 128;
	setp.ne.s32 	%p344, %r41, %r488;
	@%p344 bra 	$Lt_38_579586;
	.loc	20	68	0
	mov.f32 	%f120, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_13;
$Lt_38_579586:
	.loc	20	70	0
	mov.u32 	%r489, 192;
	setp.ne.s32 	%p345, %r41, %r489;
	@%p345 bra 	$Lt_38_579842;
	.loc	20	72	0
	mov.f32 	%f120, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_13;
$Lt_38_579842:
	.loc	20	76	0
	mov.f32 	%f120, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_13:
	.loc	20	546	0
	mov.f32 	%f10, %f120;
	bra.uni 	$Lt_38_758786;
$Lt_38_759042:
	.loc	20	433	0
	mov.f32 	%f145, %f10;
	mov.f32 	%f475, %f12;
	mov.f32 	%f147, %f475;
	mov.f32 	%f476, %f13;
	mov.f32 	%f149, %f476;
	mov.f32 	%f477, %f14;
	mov.f32 	%f151, %f477;
	.loc	20	435	0
	mov.u32 	%r490, 0;
	setp.eq.s32 	%p346, %r46, %r490;
	@%p346 bra 	$Lt_38_763906;
	.loc	20	57	0
	mov.u32 	%r491, 0;
	setp.ne.s32 	%p347, %r41, %r491;
	@%p347 bra 	$Lt_38_580354;
	.loc	20	59	0
	mov.f32 	%f152, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_11;
$Lt_38_580354:
	.loc	20	61	0
	mov.u32 	%r492, 64;
	setp.ne.s32 	%p348, %r41, %r492;
	@%p348 bra 	$Lt_38_580610;
	.loc	20	63	0
	mov.f32 	%f152, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_11;
$Lt_38_580610:
	.loc	20	65	0
	mov.u32 	%r493, 128;
	setp.ne.s32 	%p349, %r41, %r493;
	@%p349 bra 	$Lt_38_580866;
	.loc	20	68	0
	mov.f32 	%f152, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_11;
$Lt_38_580866:
	.loc	20	70	0
	mov.u32 	%r494, 192;
	setp.ne.s32 	%p350, %r41, %r494;
	@%p350 bra 	$Lt_38_581122;
	.loc	20	72	0
	mov.f32 	%f152, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_11;
$Lt_38_581122:
	.loc	20	76	0
	mov.f32 	%f152, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_11:
	.loc	20	118	0
	and.b32 	%r148, %r20, 2048;
	mov.s32 	%r495, 0;
	setp.ne.s32 	%p89, %r148, %r495;
	@!%p89 bra 	$Lt_38_764674;
	.loc	20	100	0
	ld.const.f32 	%f478, [kYCbCrOffset+0];
	bra.uni 	$Lt_38_764418;
$Lt_38_764674:
	ld.const.f32 	%f478, [kYCbCrFullRangeOffset+0];
$Lt_38_764418:
	.loc	20	118	0
	@!%p89 bra 	$Lt_38_765186;
	.loc	20	100	0
	ld.const.f32 	%f479, [kYCbCrOffset+4];
	bra.uni 	$Lt_38_764930;
$Lt_38_765186:
	ld.const.f32 	%f479, [kYCbCrFullRangeOffset+4];
$Lt_38_764930:
	.loc	20	118	0
	@!%p89 bra 	$Lt_38_765698;
	.loc	20	100	0
	ld.const.f32 	%f480, [kYCbCrOffset+8];
	bra.uni 	$Lt_38_765442;
$Lt_38_765698:
	ld.const.f32 	%f480, [kYCbCrFullRangeOffset+8];
$Lt_38_765442:
	.loc	20	437	0
	mov.f32 	%f481, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f482, %f152, %f481;
	mul.ftz.f32 	%f483, %f482, %f478;
	sub.ftz.f32 	%f147, %f475, %f483;
	mul.ftz.f32 	%f484, %f482, %f479;
	sub.ftz.f32 	%f149, %f476, %f484;
	mul.ftz.f32 	%f485, %f482, %f480;
	sub.ftz.f32 	%f151, %f477, %f485;
$Lt_38_763906:
	mov.f32 	%f486, 0fb70637bd;   	// -8e-006
	add.ftz.f32 	%f487, %f10, %f486;
	mov.f32 	%f488, 0f00000000;   	// 0
	setp.le.ftz.f32 	%p351, %f487, %f488;
	@!%p351 bra 	$Lt_38_766210;
	mov.f32 	%f151, 0f00000000;   	// 0
	mov.f32 	%f149, 0f00000000;   	// 0
	mov.f32 	%f147, 0f00000000;   	// 0
	mov.f32 	%f145, 0f00000000;   	// 0
	bra.uni 	$Lt_38_765954;
$Lt_38_766210:
	.loc	20	57	0
	mov.u32 	%r496, 0;
	setp.ne.s32 	%p352, %r41, %r496;
	@%p352 bra 	$Lt_38_583426;
	.loc	20	59	0
	mov.f32 	%f164, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_7;
$Lt_38_583426:
	.loc	20	61	0
	mov.u32 	%r497, 64;
	setp.ne.s32 	%p353, %r41, %r497;
	@%p353 bra 	$Lt_38_583682;
	.loc	20	63	0
	mov.f32 	%f164, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_7;
$Lt_38_583682:
	.loc	20	65	0
	mov.u32 	%r498, 128;
	setp.ne.s32 	%p354, %r41, %r498;
	@%p354 bra 	$Lt_38_583938;
	.loc	20	68	0
	mov.f32 	%f164, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_7;
$Lt_38_583938:
	.loc	20	70	0
	mov.u32 	%r499, 192;
	setp.ne.s32 	%p355, %r41, %r499;
	@%p355 bra 	$Lt_38_584194;
	.loc	20	72	0
	mov.f32 	%f164, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_7;
$Lt_38_584194:
	.loc	20	76	0
	mov.f32 	%f164, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_7:
	.loc	20	447	0
	div.approx.ftz.f32 	%f489, %f164, %f10;
	mul.ftz.f32 	%f147, %f489, %f147;
	.loc	20	448	0
	mul.ftz.f32 	%f149, %f489, %f149;
	.loc	20	449	0
	mul.ftz.f32 	%f151, %f489, %f151;
$Lt_38_765954:
	.loc	20	452	0
	mov.u32 	%r500, 0;
	setp.eq.s32 	%p356, %r46, %r500;
	@%p356 bra 	$Lt_38_766466;
	.loc	20	57	0
	mov.u32 	%r501, 0;
	setp.ne.s32 	%p357, %r41, %r501;
	@%p357 bra 	$Lt_38_584706;
	.loc	20	59	0
	mov.f32 	%f166, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_5;
$Lt_38_584706:
	.loc	20	61	0
	mov.u32 	%r502, 64;
	setp.ne.s32 	%p358, %r41, %r502;
	@%p358 bra 	$Lt_38_584962;
	.loc	20	63	0
	mov.f32 	%f166, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_5;
$Lt_38_584962:
	.loc	20	65	0
	mov.u32 	%r503, 128;
	setp.ne.s32 	%p359, %r41, %r503;
	@%p359 bra 	$Lt_38_585218;
	.loc	20	68	0
	mov.f32 	%f166, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_5;
$Lt_38_585218:
	.loc	20	70	0
	mov.u32 	%r504, 192;
	setp.ne.s32 	%p360, %r41, %r504;
	@%p360 bra 	$Lt_38_585474;
	.loc	20	72	0
	mov.f32 	%f166, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_5;
$Lt_38_585474:
	.loc	20	76	0
	mov.f32 	%f166, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_5:
	.loc	20	107	0
	and.b32 	%r148, %r20, 2048;
	mov.s32 	%r505, 0;
	setp.ne.s32 	%p89, %r148, %r505;
	@!%p89 bra 	$Lt_38_767234;
	.loc	20	100	0
	ld.const.f32 	%f490, [kYCbCrOffset+0];
	bra.uni 	$Lt_38_766978;
$Lt_38_767234:
	ld.const.f32 	%f490, [kYCbCrFullRangeOffset+0];
$Lt_38_766978:
	.loc	20	107	0
	@!%p89 bra 	$Lt_38_767746;
	.loc	20	100	0
	ld.const.f32 	%f491, [kYCbCrOffset+4];
	bra.uni 	$Lt_38_767490;
$Lt_38_767746:
	ld.const.f32 	%f491, [kYCbCrFullRangeOffset+4];
$Lt_38_767490:
	.loc	20	107	0
	@!%p89 bra 	$Lt_38_768258;
	.loc	20	100	0
	ld.const.f32 	%f492, [kYCbCrOffset+8];
	bra.uni 	$Lt_38_768002;
$Lt_38_768258:
	ld.const.f32 	%f492, [kYCbCrFullRangeOffset+8];
$Lt_38_768002:
	.loc	20	454	0
	mov.f32 	%f493, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f494, %f166, %f493;
	fma.rn.ftz.f32 	%f147, %f494, %f490, %f147;
	fma.rn.ftz.f32 	%f149, %f494, %f491, %f149;
	fma.rn.ftz.f32 	%f151, %f494, %f492, %f151;
$Lt_38_766466:
	.loc	20	551	0
	mov.f32 	%f10, %f145;
	mov.f32 	%f12, %f147;
	mov.f32 	%f13, %f149;
	mov.f32 	%f14, %f151;
$Lt_38_758786:
$L_38_685826:
$Lt_38_758274:
	.loc	20	540	0
	@!%p130 bra 	$Lt_38_768514;
	.loc	21	268	0
	mov.f32 	%f495, %f12;
	.loc	21	269	0
	mov.f32 	%f496, %f10;
	.loc	20	558	0
	mov.f32 	%f10, %f14;
	mov.f32 	%f12, %f13;
	mov.f32 	%f13, %f495;
	mov.f32 	%f14, %f496;
$Lt_38_768514:
	@!%p8 bra 	$Lt_38_769026;
	.loc	20	57	0
	mov.u32 	%r506, 0;
	setp.ne.s32 	%p361, %r41, %r506;
	@%p361 bra 	$Lt_38_588034;
	.loc	20	59	0
	mov.f32 	%f174, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_1;
$Lt_38_588034:
	.loc	20	61	0
	mov.u32 	%r507, 64;
	setp.ne.s32 	%p362, %r41, %r507;
	@%p362 bra 	$Lt_38_588290;
	.loc	20	63	0
	mov.f32 	%f174, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_1;
$Lt_38_588290:
	.loc	20	65	0
	mov.u32 	%r508, 128;
	setp.ne.s32 	%p363, %r41, %r508;
	@%p363 bra 	$Lt_38_588546;
	.loc	20	68	0
	mov.f32 	%f174, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_1;
$Lt_38_588546:
	.loc	20	70	0
	mov.u32 	%r509, 192;
	setp.ne.s32 	%p364, %r41, %r509;
	@%p364 bra 	$Lt_38_588802;
	.loc	20	72	0
	mov.f32 	%f174, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_216_1;
$Lt_38_588802:
	.loc	20	76	0
	mov.f32 	%f174, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_216_1:
	.loc	20	564	0
	mov.f32 	%f497, 0f3f000000;   	// 0.5
	add.ftz.f32 	%f498, %f10, %f497;
	mov.f32 	%f499, 0f00000000;   	// 0
	mov.f32 	%f500, 0f00000000;   	// 0
	setp.gt.ftz.f32 	%p365, %f498, %f500;
	selp.f32 	%f501, %f498, %f499, %p365;
	min.ftz.f32 	%f10, %f501, %f174;
	mov.f32 	%f502, 0f3f000000;   	// 0.5
	add.ftz.f32 	%f503, %f12, %f502;
	mov.f32 	%f504, 0f00000000;   	// 0
	mov.f32 	%f505, 0f00000000;   	// 0
	setp.gt.ftz.f32 	%p366, %f503, %f505;
	selp.f32 	%f506, %f503, %f504, %p366;
	min.ftz.f32 	%f12, %f506, %f174;
	mov.f32 	%f507, 0f3f000000;   	// 0.5
	add.ftz.f32 	%f508, %f13, %f507;
	mov.f32 	%f509, 0f00000000;   	// 0
	mov.f32 	%f510, 0f00000000;   	// 0
	setp.gt.ftz.f32 	%p367, %f508, %f510;
	selp.f32 	%f511, %f508, %f509, %p367;
	min.ftz.f32 	%f13, %f511, %f174;
	mov.f32 	%f512, 0f3f000000;   	// 0.5
	add.ftz.f32 	%f513, %f14, %f512;
	mov.f32 	%f514, 0f00000000;   	// 0
	mov.f32 	%f515, 0f00000000;   	// 0
	setp.gt.ftz.f32 	%p368, %f513, %f515;
	selp.f32 	%f516, %f513, %f514, %p368;
	min.ftz.f32 	%f14, %f516, %f174;
$Lt_38_769026:
	st.v4.f32 	[%rd4+0], {%f10,%f12,%f13,%f14};
$Lt_38_714754:
	.loc	22	107	0
	ret;
$LDWend__Z33ConvertPixelTwo_422_8u_To_444_32f6uchar4S_14IR_PixelFormatP6float4S2_S0_:
	} // _Z33ConvertPixelTwo_422_8u_To_444_32f6uchar4S_14IR_PixelFormatP6float4S2_S0_

	.visible .func (.param .align 16 .b8 __cudaretf__Z30ConvertPixel_422_8u_To_444_32f6uchar414IR_PixelFormatS0_[16]) _Z30ConvertPixel_422_8u_To_444_32f6uchar414IR_PixelFormatS0_ (.param .align 4 .b8 __cudaparmf1__Z30ConvertPixel_422_8u_To_444_32f6uchar414IR_PixelFormatS0_[4], .param .s32 __cudaparmf2__Z30ConvertPixel_422_8u_To_444_32f6uchar414IR_PixelFormatS0_, .param .s32 __cudaparmf3__Z30ConvertPixel_422_8u_To_444_32f6uchar414IR_PixelFormatS0_)
	{
	.reg .u32 %r<206>;
	.reg .u64 %rd<3>;
	.reg .f32 %f<197>;
	.reg .pred %p<138>;
	.loc	22	112	0
$LDWbegin__Z30ConvertPixel_422_8u_To_444_32f6uchar414IR_PixelFormatS0_:
	ld.param.u8 	%r1, [__cudaparmf1__Z30ConvertPixel_422_8u_To_444_32f6uchar414IR_PixelFormatS0_+0];
	mov.s32 	%r2, %r1;
	ld.param.u8 	%r3, [__cudaparmf1__Z30ConvertPixel_422_8u_To_444_32f6uchar414IR_PixelFormatS0_+1];
	mov.s32 	%r4, %r3;
	ld.param.u8 	%r5, [__cudaparmf1__Z30ConvertPixel_422_8u_To_444_32f6uchar414IR_PixelFormatS0_+2];
	mov.s32 	%r6, %r5;
	ld.param.u8 	%r7, [__cudaparmf1__Z30ConvertPixel_422_8u_To_444_32f6uchar414IR_PixelFormatS0_+3];
	mov.s32 	%r8, %r7;
	ld.param.u32 	%r9, [__cudaparmf2__Z30ConvertPixel_422_8u_To_444_32f6uchar414IR_PixelFormatS0_];
	mov.s32 	%r10, %r9;
	ld.param.u32 	%r11, [__cudaparmf3__Z30ConvertPixel_422_8u_To_444_32f6uchar414IR_PixelFormatS0_];
	mov.s32 	%r12, %r11;
	.loc	22	113	0
	cvt.u8.u32 	%r13, %r2;
	cvt.rn.f32.u32 	%f1, %r13;
	mov.f32 	%f2, %f1;
	cvt.u8.u32 	%r14, %r4;
	cvt.rn.f32.u32 	%f3, %r14;
	cvt.u8.u32 	%r15, %r6;
	cvt.rn.f32.u32 	%f4, %r15;
	and.b32 	%r16, %r10, 4096;
	mov.u32 	%r17, 0;
	setp.eq.s32 	%p1, %r16, %r17;
	@%p1 bra 	$Lt_39_228866;
	.loc	23	86	0
	mov.f32 	%f2, %f3;
	.loc	23	87	0
	mov.f32 	%f3, %f1;
	.loc	22	38	0
	cvt.u8.u32 	%r18, %r8;
	cvt.rn.f32.u32 	%f4, %r18;
$Lt_39_228866:
	.loc	20	57	0
	and.b32 	%r19, %r10, 1536;
	and.b32 	%r20, %r10, 448;
	mov.s32 	%r21, 512;
	setp.eq.s32 	%p2, %r19, %r21;
	mov.s32 	%r22, 2565;
	mov.s32 	%r23, 2053;
	selp.s32 	%r24, %r22, %r23, %p2;
	or.b32 	%r25, %r24, %r20;
	and.b32 	%r26, %r25, 448;
	mov.s32 	%r27, 0;
	setp.eq.s32 	%p3, %r26, %r27;
	@!%p3 bra 	$Lt_39_144642;
	.loc	20	59	0
	mov.f32 	%f5, 0f437f0000;     	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_77;
$Lt_39_144642:
	.loc	20	61	0
	mov.u32 	%r28, 64;
	setp.ne.s32 	%p4, %r26, %r28;
	@%p4 bra 	$Lt_39_144898;
	.loc	20	63	0
	mov.f32 	%f5, 0f447fc000;     	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_77;
$Lt_39_144898:
	.loc	20	65	0
	mov.u32 	%r29, 128;
	setp.ne.s32 	%p5, %r26, %r29;
	@%p5 bra 	$Lt_39_145154;
	.loc	20	68	0
	mov.f32 	%f5, 0f47000000;     	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_77;
$Lt_39_145154:
	.loc	20	70	0
	mov.u32 	%r30, 192;
	setp.ne.s32 	%p6, %r26, %r30;
	@%p6 bra 	$Lt_39_145410;
	.loc	20	72	0
	mov.f32 	%f5, 0fbf800000;     	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_77;
$Lt_39_145410:
	.loc	20	76	0
	mov.f32 	%f5, 0f3f800000;     	// 1
$LDWendi__Z19MaxUnsignedBitValuei_217_77:
	.loc	20	469	0
	mov.f32 	%f6, %f4;
	mov.f32 	%f7, %f6;
	mov.f32 	%f8, %f2;
	mov.f32 	%f9, %f8;
	mov.f32 	%f10, %f3;
	mov.f32 	%f11, %f5;
	and.b32 	%r31, %r25, 4096;
	mov.u32 	%r32, 0;
	setp.ne.s32 	%p7, %r31, %r32;
	@%p7 bra 	$Lt_39_229378;
	.loc	20	473	0
	mov.f32 	%f7, %f11;
	mov.f32 	%f9, %f10;
	mov.f32 	%f10, %f8;
	mov.f32 	%f11, %f6;
$Lt_39_229378:
	.loc	20	476	0
	and.b32 	%r33, %r12, 448;
	mov.s32 	%r34, %r12;
	mov.s32 	%r35, 256;
	setp.ne.s32 	%p8, %r33, %r35;
	and.b32 	%r36, %r34, 1;
	selp.s32 	%r37, 1, 0, %p8;
	and.b32 	%r38, %r36, %r37;
	mov.s32 	%r39, %r25;
	mov.s32 	%r40, 256;
	setp.ne.s32 	%p9, %r26, %r40;
	and.b32 	%r41, %r39, 1;
	selp.s32 	%r42, 1, 0, %p9;
	and.b32 	%r43, %r41, %r42;
	mov.u32 	%r44, 0;
	setp.eq.s32 	%p10, %r43, %r44;
	@%p10 bra 	$Lt_39_256514;
	mov.u32 	%r45, 0;
	setp.ne.s32 	%p11, %r38, %r45;
	@%p11 bra 	$Lt_39_256514;
	.loc	20	57	0
	@!%p3 bra 	$Lt_39_145922;
	.loc	20	59	0
	mov.f32 	%f12, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_75;
$Lt_39_145922:
	.loc	20	61	0
	mov.u32 	%r46, 64;
	setp.ne.s32 	%p12, %r26, %r46;
	@%p12 bra 	$Lt_39_146178;
	.loc	20	63	0
	mov.f32 	%f12, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_75;
$Lt_39_146178:
	.loc	20	65	0
	mov.u32 	%r47, 128;
	setp.ne.s32 	%p13, %r26, %r47;
	@%p13 bra 	$Lt_39_146434;
	.loc	20	68	0
	mov.f32 	%f12, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_75;
$Lt_39_146434:
	.loc	20	70	0
	mov.u32 	%r48, 192;
	setp.ne.s32 	%p14, %r26, %r48;
	@%p14 bra 	$Lt_39_146690;
	.loc	20	72	0
	mov.f32 	%f12, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_75;
$Lt_39_146690:
	.loc	20	76	0
	mov.f32 	%f12, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_217_75:
	.loc	20	118	0
	and.b32 	%r49, %r25, 2048;
	mov.s32 	%r50, 0;
	setp.ne.s32 	%p15, %r49, %r50;
	@!%p15 bra 	$Lt_39_230146;
	.loc	20	100	0
	ld.const.f32 	%f13, [kYCbCrOffset+0];
	bra.uni 	$Lt_39_229890;
$Lt_39_230146:
	ld.const.f32 	%f13, [kYCbCrFullRangeOffset+0];
$Lt_39_229890:
	.loc	20	118	0
	@!%p15 bra 	$Lt_39_230658;
	.loc	20	100	0
	ld.const.f32 	%f14, [kYCbCrOffset+4];
	bra.uni 	$Lt_39_230402;
$Lt_39_230658:
	ld.const.f32 	%f14, [kYCbCrFullRangeOffset+4];
$Lt_39_230402:
	.loc	20	118	0
	@!%p15 bra 	$Lt_39_231170;
	.loc	20	100	0
	ld.const.f32 	%f15, [kYCbCrOffset+8];
	bra.uni 	$Lt_39_230914;
$Lt_39_231170:
	ld.const.f32 	%f15, [kYCbCrFullRangeOffset+8];
$Lt_39_230914:
	.loc	20	478	0
	mov.f32 	%f16, 0f437f0000;    	// 255
	div.approx.ftz.f32 	%f17, %f12, %f16;
	mul.ftz.f32 	%f18, %f17, %f13;
	sub.ftz.f32 	%f9, %f9, %f18;
	mul.ftz.f32 	%f19, %f17, %f14;
	sub.ftz.f32 	%f10, %f10, %f19;
	mul.ftz.f32 	%f20, %f17, %f15;
	sub.ftz.f32 	%f11, %f11, %f20;
$Lt_39_256514:
$Lt_39_29698:
	.loc	20	481	0
	and.b32 	%r51, %r12, 2;
	and.b32 	%r52, %r25, 2;
	mov.u32 	%r53, 0;
	setp.eq.s32 	%p16, %r52, %r53;
	@%p16 bra 	$Lt_39_257026;
	mov.u32 	%r54, 0;
	setp.ne.s32 	%p17, %r51, %r54;
	@%p17 bra 	$Lt_39_257026;
	.loc	20	483	0
	mov.f32 	%f21, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p18, %f9, %f21;
	@!%p18 bra 	$Lt_39_231426;
	.loc	20	372	0
	neg.ftz.f32 	%f22, %f9;
	lg2.approx.ftz.f32 	%f23, %f22;
	mov.f32 	%f24, 0f3ee66666;    	// 0.45
	mul.ftz.f32 	%f25, %f23, %f24;
	ex2.approx.ftz.f32 	%f26, %f25;
	neg.ftz.f32 	%f27, %f26;
	bra.uni 	$LDWendi___log2f_217_71;
$Lt_39_231426:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f28, %f9;
	mov.f32 	%f29, 0f3ee66666;    	// 0.45
	mul.ftz.f32 	%f30, %f28, %f29;
	ex2.approx.ftz.f32 	%f27, %f30;
$LDWendi___log2f_217_71:
	.loc	20	483	0
	mov.f32 	%f31, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p19, %f10, %f31;
	@!%p19 bra 	$Lt_39_231938;
	.loc	20	372	0
	neg.ftz.f32 	%f32, %f10;
	lg2.approx.ftz.f32 	%f33, %f32;
	mov.f32 	%f34, 0f3ee66666;    	// 0.45
	mul.ftz.f32 	%f35, %f33, %f34;
	ex2.approx.ftz.f32 	%f36, %f35;
	neg.ftz.f32 	%f37, %f36;
	bra.uni 	$LDWendi___log2f_217_69;
$Lt_39_231938:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f38, %f10;
	mov.f32 	%f39, 0f3ee66666;    	// 0.45
	mul.ftz.f32 	%f40, %f38, %f39;
	ex2.approx.ftz.f32 	%f37, %f40;
$LDWendi___log2f_217_69:
	.loc	20	483	0
	mov.f32 	%f41, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p20, %f11, %f41;
	@!%p20 bra 	$Lt_39_232450;
	.loc	20	372	0
	neg.ftz.f32 	%f42, %f11;
	lg2.approx.ftz.f32 	%f43, %f42;
	mov.f32 	%f44, 0f3ee66666;    	// 0.45
	mul.ftz.f32 	%f45, %f43, %f44;
	ex2.approx.ftz.f32 	%f46, %f45;
	neg.ftz.f32 	%f47, %f46;
	bra.uni 	$LDWendi___log2f_217_67;
$Lt_39_232450:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f48, %f11;
	mov.f32 	%f49, 0f3ee66666;    	// 0.45
	mul.ftz.f32 	%f50, %f48, %f49;
	ex2.approx.ftz.f32 	%f47, %f50;
$LDWendi___log2f_217_67:
	.loc	20	483	0
	mov.f32 	%f9, %f27;
	mov.f32 	%f10, %f37;
	mov.f32 	%f11, %f47;
$Lt_39_257026:
$Lt_39_33538:
	.loc	20	486	0
	and.b32 	%r55, %r12, 1;
	and.b32 	%r56, %r12, 1536;
	and.b32 	%r57, %r25, 1;
	and.b32 	%r58, %r25, 1536;
	set.ne.u32.s32 	%r59, %r55, %r57;
	neg.s32 	%r60, %r59;
	set.ne.u32.s32 	%r61, %r56, %r58;
	neg.s32 	%r62, %r61;
	or.b32 	%r63, %r60, %r62;
	mov.u32 	%r64, 0;
	setp.ne.s32 	%p21, %r63, %r64;
	@%p21 bra 	$Lt_39_84226;
	setp.eq.s32 	%p22, %r38, %r43;
	@%p22 bra 	$Lt_39_84482;
$Lt_39_84226:
	.loc	20	490	0
	mov.u32 	%r65, 0;
	setp.ne.s32 	%p23, %r57, %r65;
	@%p23 bra 	$Lt_39_233218;
	mov.s32 	%r66, 256;
	setp.eq.s32 	%p24, %r33, %r66;
	mov.u32 	%r67, 256;
	setp.ne.s32 	%p25, %r26, %r67;
	@%p25 bra 	$Lt_39_233730;
	.loc	20	137	0
	mov.s32 	%r68, 512;
	setp.eq.s32 	%p26, %r56, %r68;
	@!%p24 bra 	$Lt_39_150018;
	.loc	20	139	0
	@!%p26 bra 	$Lt_39_150274;
	.loc	20	141	0
	mov.u64 	%rd1, kRGB32f_To_709YPbPr;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__217_65;
$Lt_39_150274:
	.loc	20	145	0
	mov.u64 	%rd1, kRGB32f_To_601YPbPr;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__217_65;
$Lt_39_150018:
	.loc	20	150	0
	@!%p26 bra 	$Lt_39_150530;
	.loc	20	152	0
	mov.u64 	%rd1, kRGB32f_To_709YCbCr;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__217_65;
$Lt_39_150530:
	.loc	20	154	0
	and.b32 	%r69, %r12, 2048;
	mov.u32 	%r70, 0;
	setp.ne.s32 	%p27, %r69, %r70;
	@%p27 bra 	$Lt_39_150786;
	.loc	20	156	0
	mov.u64 	%rd1, kRGB32f_To_601YCbCrFullRange;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__217_65;
$Lt_39_150786:
	.loc	20	160	0
	mov.u64 	%rd1, kRGB32f_To_601YCbCr;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__217_65;
$Lt_39_233730:
	@!%p24 bra 	$Lt_39_234242;
	bra.uni 	$Lt_39_232962;
$Lt_39_234242:
	.loc	20	179	0
	mov.u32 	%r71, 512;
	setp.ne.s32 	%p28, %r56, %r71;
	@%p28 bra 	$Lt_39_151554;
	.loc	20	181	0
	mov.u64 	%rd1, kRGB8u_To_709YCbCr;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__217_65;
$Lt_39_151554:
	.loc	20	183	0
	and.b32 	%r72, %r12, 2048;
	mov.u32 	%r73, 0;
	setp.ne.s32 	%p29, %r72, %r73;
	@%p29 bra 	$Lt_39_151810;
	.loc	20	185	0
	mov.u64 	%rd1, kRGB8u_To_601YCbCrFullRange;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__217_65;
$Lt_39_151810:
	.loc	20	189	0
	mov.u64 	%rd1, kRGB8u_To_601YCbCr;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__217_65;
$Lt_39_233218:
	mov.s32 	%r74, 0;
	setp.eq.s32 	%p30, %r55, %r74;
	mov.u32 	%r75, 512;
	setp.ne.s32 	%p31, %r58, %r75;
	@%p31 bra 	$Lt_39_234754;
	mov.s32 	%r76, 256;
	setp.eq.s32 	%p32, %r26, %r76;
	@!%p30 bra 	$Lt_39_235266;
	mov.s32 	%r77, 256;
	setp.eq.s32 	%p33, %r33, %r77;
	@!%p32 bra 	$Lt_39_235778;
	@!%p33 bra 	$Lt_39_232962;
	.loc	20	202	0
	mov.u64 	%rd1, k709YPbPr_To_RGB32f;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__217_65;
$Lt_39_235778:
	.loc	20	211	0
	@!%p33 bra 	$Lt_39_153090;
	.loc	20	213	0
	mov.u64 	%rd1, k709YCbCr_To_RGB32f;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__217_65;
$Lt_39_153090:
	.loc	20	217	0
	mov.u64 	%rd1, k709YCbCr_To_RGB8u;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__217_65;
$Lt_39_235266:
	@!%p32 bra 	$Lt_39_236802;
	bra.uni 	$Lt_39_232962;
$Lt_39_236802:
	mov.s32 	%r78, 256;
	set.eq.u32.s32 	%r79, %r33, %r78;
	neg.s32 	%r80, %r79;
	and.b32 	%r81, %r12, 2048;
	mov.s32 	%r82, 0;
	set.eq.u32.s32 	%r83, %r81, %r82;
	neg.s32 	%r84, %r83;
	or.b32 	%r85, %r80, %r84;
	mov.u32 	%r86, 0;
	setp.eq.s32 	%p34, %r85, %r86;
	@%p34 bra 	$Lt_39_237314;
	bra.uni 	$Lt_39_232962;
$Lt_39_237314:
	.loc	20	250	0
	mov.u64 	%rd1, k709YCbCr_To_601YCbCr;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__217_65;
$Lt_39_234754:
	and.b32 	%r87, %r25, 2048;
	mov.s32 	%r88, 0;
	setp.eq.s32 	%p35, %r87, %r88;
	@!%p35 bra 	$Lt_39_237826;
	@!%p30 bra 	$Lt_39_232962;
	.loc	20	259	0
	mov.u32 	%r89, 256;
	setp.ne.s32 	%p36, %r33, %r89;
	@%p36 bra 	$Lt_39_154626;
	.loc	20	261	0
	mov.u64 	%rd1, k601YCbCrFullRange_To_RGB32f;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__217_65;
$Lt_39_154626:
	.loc	20	265	0
	mov.u64 	%rd1, k601YCbCrFullRange_To_RGB8u;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__217_65;
$Lt_39_237826:
	mov.s32 	%r90, 256;
	setp.eq.s32 	%p32, %r26, %r90;
	@!%p30 bra 	$Lt_39_238850;
	mov.s32 	%r91, 256;
	setp.eq.s32 	%p37, %r33, %r91;
	@!%p32 bra 	$Lt_39_239362;
	@!%p37 bra 	$Lt_39_232962;
	.loc	20	302	0
	mov.u64 	%rd1, k601YPbPr_To_RGB32f;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__217_65;
$Lt_39_239362:
	.loc	20	311	0
	@!%p37 bra 	$Lt_39_156418;
	.loc	20	313	0
	mov.u64 	%rd1, k601YCbCr_To_RGB32f;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__217_65;
$Lt_39_156418:
	.loc	20	317	0
	mov.u64 	%rd1, k601YCbCr_To_RGB8u;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__217_65;
$Lt_39_238850:
	@!%p32 bra 	$Lt_39_240386;
	bra.uni 	$Lt_39_232962;
$Lt_39_240386:
	selp.s32 	%r92, 1, 0, %p35;
	mov.s32 	%r93, 256;
	set.eq.u32.s32 	%r94, %r33, %r93;
	neg.s32 	%r95, %r94;
	or.b32 	%r96, %r92, %r95;
	mov.u32 	%r97, 0;
	setp.eq.s32 	%p38, %r96, %r97;
	@%p38 bra 	$Lt_39_240898;
	bra.uni 	$Lt_39_232962;
$Lt_39_240898:
	.loc	20	350	0
	mov.u64 	%rd1, k601YCbCr_To_709YCbCr;
	bra.uni 	$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__217_65;
$Lt_39_232962:
	.loc	20	355	0
	mov.u64 	%rd1, 0;
$LDWendi__Z23ColorSpaceConvertMatrix14IR_PixelFormatS__217_65:
	.loc	20	490	0
	ld.global.f32 	%f51, [%rd1+16];
	mul.ftz.f32 	%f52, %f51, %f10;
	ld.global.f32 	%f53, [%rd1+12];
	fma.rn.ftz.f32 	%f54, %f53, %f9, %f52;
	ld.global.f32 	%f55, [%rd1+20];
	fma.rn.ftz.f32 	%f56, %f55, %f11, %f54;
	ld.global.f32 	%f57, [%rd1+28];
	mul.ftz.f32 	%f58, %f57, %f10;
	ld.global.f32 	%f59, [%rd1+24];
	fma.rn.ftz.f32 	%f60, %f59, %f9, %f58;
	ld.global.f32 	%f61, [%rd1+32];
	fma.rn.ftz.f32 	%f62, %f61, %f11, %f60;
	ld.global.f32 	%f63, [%rd1+4];
	mul.ftz.f32 	%f64, %f63, %f10;
	ld.global.f32 	%f65, [%rd1+0];
	fma.rn.ftz.f32 	%f66, %f65, %f9, %f64;
	ld.global.f32 	%f67, [%rd1+8];
	fma.rn.ftz.f32 	%f9, %f67, %f11, %f66;
	mov.f32 	%f10, %f56;
	mov.f32 	%f11, %f62;
	setp.eq.s32 	%p39, %r33, %r26;
	@%p39 bra 	$Lt_39_241666;
	.loc	20	494	0
	mov.s32 	%r98, 256;
	setp.eq.s32 	%p32, %r26, %r98;
	@!%p32 bra 	$L_39_226306;
	mov.s32 	%r99, 0;
	setp.eq.s32 	%p40, %r33, %r99;
	@%p40 bra 	$L_39_226050;
$L_39_226306:
	@!%p3 bra 	$Lt_39_258306;
	mov.u32 	%r100, 256;
	setp.ne.s32 	%p41, %r33, %r100;
	@%p41 bra 	$Lt_39_258306;
	mov.s32 	%r101, 0;
	setp.eq.s32 	%p40, %r33, %r101;
$L_39_226050:
	.loc	20	57	0
	@!%p40 bra 	$Lt_39_157954;
	.loc	20	59	0
	mov.f32 	%f68, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_63;
$Lt_39_157954:
	.loc	20	61	0
	mov.u32 	%r102, 64;
	setp.ne.s32 	%p42, %r33, %r102;
	@%p42 bra 	$Lt_39_158210;
	.loc	20	63	0
	mov.f32 	%f68, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_63;
$Lt_39_158210:
	.loc	20	65	0
	mov.u32 	%r103, 128;
	setp.ne.s32 	%p43, %r33, %r103;
	@%p43 bra 	$Lt_39_158466;
	.loc	20	68	0
	mov.f32 	%f68, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_63;
$Lt_39_158466:
	.loc	20	70	0
	mov.u32 	%r104, 192;
	setp.ne.s32 	%p44, %r33, %r104;
	@%p44 bra 	$Lt_39_158722;
	.loc	20	72	0
	mov.f32 	%f68, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_63;
$Lt_39_158722:
	.loc	20	76	0
	mov.f32 	%f68, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_217_63:
	.loc	20	57	0
	@!%p3 bra 	$Lt_39_158978;
	.loc	20	59	0
	mov.f32 	%f69, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_61;
$Lt_39_158978:
	.loc	20	61	0
	mov.u32 	%r105, 64;
	setp.ne.s32 	%p45, %r26, %r105;
	@%p45 bra 	$Lt_39_159234;
	.loc	20	63	0
	mov.f32 	%f69, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_61;
$Lt_39_159234:
	.loc	20	65	0
	mov.u32 	%r106, 128;
	setp.ne.s32 	%p46, %r26, %r106;
	@%p46 bra 	$Lt_39_159490;
	.loc	20	68	0
	mov.f32 	%f69, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_61;
$Lt_39_159490:
	.loc	20	70	0
	mov.u32 	%r107, 192;
	setp.ne.s32 	%p47, %r26, %r107;
	@%p47 bra 	$Lt_39_159746;
	.loc	20	72	0
	mov.f32 	%f69, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_61;
$Lt_39_159746:
	.loc	20	76	0
	mov.f32 	%f69, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_217_61:
	.loc	20	498	0
	div.approx.ftz.f32 	%f70, %f68, %f69;
	mul.ftz.f32 	%f7, %f7, %f70;
	bra.uni 	$Lt_39_241666;
$Lt_39_258306:
$L_39_225794:
	.loc	20	500	0
	@!%p32 bra 	$L_39_227330;
	@%p8 bra 	$L_39_227074;
$L_39_227330:
	@!%p9 bra 	$Lt_39_259330;
	mov.u32 	%r108, 256;
	setp.ne.s32 	%p48, %r33, %r108;
	@%p48 bra 	$Lt_39_259330;
$L_39_227074:
	.loc	20	57	0
	mov.u32 	%r109, 0;
	setp.ne.s32 	%p49, %r33, %r109;
	@%p49 bra 	$Lt_39_160258;
	.loc	20	59	0
	mov.f32 	%f68, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_59;
$Lt_39_160258:
	.loc	20	61	0
	mov.u32 	%r110, 64;
	setp.ne.s32 	%p50, %r33, %r110;
	@%p50 bra 	$Lt_39_160514;
	.loc	20	63	0
	mov.f32 	%f68, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_59;
$Lt_39_160514:
	.loc	20	65	0
	mov.u32 	%r111, 128;
	setp.ne.s32 	%p51, %r33, %r111;
	@%p51 bra 	$Lt_39_160770;
	.loc	20	68	0
	mov.f32 	%f68, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_59;
$Lt_39_160770:
	.loc	20	70	0
	mov.u32 	%r112, 192;
	setp.ne.s32 	%p52, %r33, %r112;
	@%p52 bra 	$Lt_39_161026;
	.loc	20	72	0
	mov.f32 	%f68, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_59;
$Lt_39_161026:
	.loc	20	76	0
	mov.f32 	%f68, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_217_59:
	.loc	20	57	0
	@!%p3 bra 	$Lt_39_161282;
	.loc	20	59	0
	mov.f32 	%f69, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_57;
$Lt_39_161282:
	.loc	20	61	0
	mov.u32 	%r113, 64;
	setp.ne.s32 	%p53, %r26, %r113;
	@%p53 bra 	$Lt_39_161538;
	.loc	20	63	0
	mov.f32 	%f69, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_57;
$Lt_39_161538:
	.loc	20	65	0
	mov.u32 	%r114, 128;
	setp.ne.s32 	%p54, %r26, %r114;
	@%p54 bra 	$Lt_39_161794;
	.loc	20	68	0
	mov.f32 	%f69, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_57;
$Lt_39_161794:
	.loc	20	70	0
	mov.u32 	%r115, 192;
	setp.ne.s32 	%p55, %r26, %r115;
	@%p55 bra 	$Lt_39_162050;
	.loc	20	72	0
	mov.f32 	%f69, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_57;
$Lt_39_162050:
	.loc	20	76	0
	mov.f32 	%f69, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_217_57:
	.loc	20	504	0
	div.approx.ftz.f32 	%f71, %f68, %f69;
	mul.ftz.f32 	%f7, %f7, %f71;
	.loc	20	57	0
	@!%p3 bra 	$Lt_39_163330;
	.loc	20	59	0
	mov.f32 	%f69, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_53;
$Lt_39_163330:
	.loc	20	61	0
	mov.u32 	%r116, 64;
	setp.ne.s32 	%p56, %r26, %r116;
	@%p56 bra 	$Lt_39_163586;
	.loc	20	63	0
	mov.f32 	%f69, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_53;
$Lt_39_163586:
	.loc	20	65	0
	mov.u32 	%r117, 128;
	setp.ne.s32 	%p57, %r26, %r117;
	@%p57 bra 	$Lt_39_163842;
	.loc	20	68	0
	mov.f32 	%f69, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_53;
$Lt_39_163842:
	.loc	20	70	0
	mov.u32 	%r118, 192;
	setp.ne.s32 	%p58, %r26, %r118;
	@%p58 bra 	$Lt_39_164098;
	.loc	20	72	0
	mov.f32 	%f69, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_53;
$Lt_39_164098:
	.loc	20	76	0
	mov.f32 	%f69, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_217_53:
	.loc	20	505	0
	mov.f32 	%f72, 0f437f0000;    	// 255
	div.approx.ftz.f32 	%f73, %f72, %f69;
	mul.ftz.f32 	%f9, %f9, %f73;
	.loc	20	57	0
	@!%p3 bra 	$Lt_39_165378;
	.loc	20	59	0
	mov.f32 	%f69, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_49;
$Lt_39_165378:
	.loc	20	61	0
	mov.u32 	%r119, 64;
	setp.ne.s32 	%p59, %r26, %r119;
	@%p59 bra 	$Lt_39_165634;
	.loc	20	63	0
	mov.f32 	%f69, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_49;
$Lt_39_165634:
	.loc	20	65	0
	mov.u32 	%r120, 128;
	setp.ne.s32 	%p60, %r26, %r120;
	@%p60 bra 	$Lt_39_165890;
	.loc	20	68	0
	mov.f32 	%f69, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_49;
$Lt_39_165890:
	.loc	20	70	0
	mov.u32 	%r121, 192;
	setp.ne.s32 	%p61, %r26, %r121;
	@%p61 bra 	$Lt_39_166146;
	.loc	20	72	0
	mov.f32 	%f69, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_49;
$Lt_39_166146:
	.loc	20	76	0
	mov.f32 	%f69, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_217_49:
	.loc	20	506	0
	mov.f32 	%f74, 0f437f0000;    	// 255
	div.approx.ftz.f32 	%f75, %f74, %f69;
	mul.ftz.f32 	%f10, %f56, %f75;
	.loc	20	57	0
	@!%p3 bra 	$Lt_39_167426;
	.loc	20	59	0
	mov.f32 	%f69, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_45;
$Lt_39_167426:
	.loc	20	61	0
	mov.u32 	%r122, 64;
	setp.ne.s32 	%p62, %r26, %r122;
	@%p62 bra 	$Lt_39_167682;
	.loc	20	63	0
	mov.f32 	%f69, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_45;
$Lt_39_167682:
	.loc	20	65	0
	mov.u32 	%r123, 128;
	setp.ne.s32 	%p63, %r26, %r123;
	@%p63 bra 	$Lt_39_167938;
	.loc	20	68	0
	mov.f32 	%f69, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_45;
$Lt_39_167938:
	.loc	20	70	0
	mov.u32 	%r124, 192;
	setp.ne.s32 	%p64, %r26, %r124;
	@%p64 bra 	$Lt_39_168194;
	.loc	20	72	0
	mov.f32 	%f69, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_45;
$Lt_39_168194:
	.loc	20	76	0
	mov.f32 	%f69, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_217_45:
	.loc	20	507	0
	mov.f32 	%f76, 0f437f0000;    	// 255
	div.approx.ftz.f32 	%f77, %f76, %f69;
	mul.ftz.f32 	%f11, %f62, %f77;
	bra.uni 	$Lt_39_241666;
$Lt_39_259330:
$L_39_226818:
	.loc	20	57	0
	@!%p3 bra 	$Lt_39_168450;
	.loc	20	59	0
	mov.f32 	%f68, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_43;
$Lt_39_168450:
	.loc	20	61	0
	mov.u32 	%r125, 64;
	setp.ne.s32 	%p65, %r26, %r125;
	@%p65 bra 	$Lt_39_168706;
	.loc	20	63	0
	mov.f32 	%f68, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_43;
$Lt_39_168706:
	.loc	20	65	0
	mov.u32 	%r126, 128;
	setp.ne.s32 	%p66, %r26, %r126;
	@%p66 bra 	$Lt_39_168962;
	.loc	20	68	0
	mov.f32 	%f68, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_43;
$Lt_39_168962:
	.loc	20	70	0
	mov.u32 	%r127, 192;
	setp.ne.s32 	%p67, %r26, %r127;
	@%p67 bra 	$Lt_39_169218;
	.loc	20	72	0
	mov.f32 	%f68, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_43;
$Lt_39_169218:
	.loc	20	76	0
	mov.f32 	%f68, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_217_43:
	.loc	20	511	0
	mov.f32 	%f78, 0f437f0000;    	// 255
	div.approx.ftz.f32 	%f79, %f68, %f78;
	mul.ftz.f32 	%f7, %f79, %f7;
	mul.ftz.f32 	%f9, %f79, %f9;
	mul.ftz.f32 	%f10, %f79, %f56;
	mul.ftz.f32 	%f11, %f79, %f62;
	bra.uni 	$Lt_39_241666;
$Lt_39_84482:
	.loc	20	486	0
	setp.eq.s32 	%p68, %r33, %r26;
	@%p68 bra 	$Lt_39_241666;
	.loc	20	57	0
	mov.u32 	%r128, 0;
	setp.ne.s32 	%p69, %r33, %r128;
	@%p69 bra 	$Lt_39_170754;
	.loc	20	59	0
	mov.f32 	%f68, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_39;
$Lt_39_170754:
	.loc	20	61	0
	mov.u32 	%r129, 64;
	setp.ne.s32 	%p70, %r33, %r129;
	@%p70 bra 	$Lt_39_171010;
	.loc	20	63	0
	mov.f32 	%f68, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_39;
$Lt_39_171010:
	.loc	20	65	0
	mov.u32 	%r130, 128;
	setp.ne.s32 	%p71, %r33, %r130;
	@%p71 bra 	$Lt_39_171266;
	.loc	20	68	0
	mov.f32 	%f68, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_39;
$Lt_39_171266:
	.loc	20	70	0
	mov.u32 	%r131, 192;
	setp.ne.s32 	%p72, %r33, %r131;
	@%p72 bra 	$Lt_39_171522;
	.loc	20	72	0
	mov.f32 	%f68, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_39;
$Lt_39_171522:
	.loc	20	76	0
	mov.f32 	%f68, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_217_39:
	.loc	20	57	0
	@!%p3 bra 	$Lt_39_171778;
	.loc	20	59	0
	mov.f32 	%f69, 0f437f0000;    	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_37;
$Lt_39_171778:
	.loc	20	61	0
	mov.u32 	%r132, 64;
	setp.ne.s32 	%p73, %r26, %r132;
	@%p73 bra 	$Lt_39_172034;
	.loc	20	63	0
	mov.f32 	%f69, 0f447fc000;    	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_37;
$Lt_39_172034:
	.loc	20	65	0
	mov.u32 	%r133, 128;
	setp.ne.s32 	%p74, %r26, %r133;
	@%p74 bra 	$Lt_39_172290;
	.loc	20	68	0
	mov.f32 	%f69, 0f47000000;    	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_37;
$Lt_39_172290:
	.loc	20	70	0
	mov.u32 	%r134, 192;
	setp.ne.s32 	%p75, %r26, %r134;
	@%p75 bra 	$Lt_39_172546;
	.loc	20	72	0
	mov.f32 	%f69, 0fbf800000;    	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_37;
$Lt_39_172546:
	.loc	20	76	0
	mov.f32 	%f69, 0f3f800000;    	// 1
$LDWendi__Z19MaxUnsignedBitValuei_217_37:
	.loc	20	517	0
	div.approx.ftz.f32 	%f80, %f68, %f69;
	mul.ftz.f32 	%f7, %f80, %f7;
	mul.ftz.f32 	%f9, %f80, %f9;
	mul.ftz.f32 	%f10, %f80, %f10;
	mul.ftz.f32 	%f11, %f80, %f11;
$Lt_39_241666:
$Lt_39_86786:
	.loc	20	520	0
	mov.u32 	%r135, 0;
	setp.eq.s32 	%p76, %r51, %r135;
	@%p76 bra 	$Lt_39_259842;
	mov.u32 	%r136, 0;
	setp.ne.s32 	%p77, %r52, %r136;
	@%p77 bra 	$Lt_39_259842;
	.loc	20	522	0
	mov.f32 	%f81, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p78, %f9, %f81;
	@!%p78 bra 	$Lt_39_242178;
	.loc	20	372	0
	neg.ftz.f32 	%f82, %f9;
	lg2.approx.ftz.f32 	%f83, %f82;
	mov.f32 	%f84, 0f400e38e4;    	// 2.22222
	mul.ftz.f32 	%f85, %f83, %f84;
	ex2.approx.ftz.f32 	%f86, %f85;
	neg.ftz.f32 	%f87, %f86;
	bra.uni 	$LDWendi___log2f_217_35;
$Lt_39_242178:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f88, %f9;
	mov.f32 	%f89, 0f400e38e4;    	// 2.22222
	mul.ftz.f32 	%f90, %f88, %f89;
	ex2.approx.ftz.f32 	%f87, %f90;
$LDWendi___log2f_217_35:
	.loc	20	522	0
	mov.f32 	%f91, 0f00000000;    	// 0
	setp.lt.ftz.f32 	%p79, %f10, %f91;
	@!%p79 bra 	$Lt_39_242690;
	.loc	20	372	0
	neg.ftz.f32 	%f92, %f10;
	lg2.approx.ftz.f32 	%f93, %f92;
	mov.f32 	%f94, 0f400e38e4;    	// 2.22222
	mul.ftz.f32 	%f95, %f93, %f94;
	ex2.approx.ftz.f32 	%f96, %f95;
	neg.ftz.f32 	%f97, %f96;
	bra.uni 	$LDWendi___log2f_217_33;
$Lt_39_242690:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f98, %f10;
	mov.f32 	%f99, 0f400e38e4;    	// 2.22222
	mul.ftz.f32 	%f100, %f98, %f99;
	ex2.approx.ftz.f32 	%f97, %f100;
$LDWendi___log2f_217_33:
	.loc	20	522	0
	mov.f32 	%f101, 0f00000000;   	// 0
	setp.lt.ftz.f32 	%p80, %f11, %f101;
	@!%p80 bra 	$Lt_39_243202;
	.loc	20	372	0
	neg.ftz.f32 	%f102, %f11;
	lg2.approx.ftz.f32 	%f103, %f102;
	mov.f32 	%f104, 0f400e38e4;   	// 2.22222
	mul.ftz.f32 	%f105, %f103, %f104;
	ex2.approx.ftz.f32 	%f106, %f105;
	neg.ftz.f32 	%f107, %f106;
	bra.uni 	$LDWendi___log2f_217_31;
$Lt_39_243202:
	.loc	20	374	0
	lg2.approx.ftz.f32 	%f108, %f11;
	mov.f32 	%f109, 0f400e38e4;   	// 2.22222
	mul.ftz.f32 	%f110, %f108, %f109;
	ex2.approx.ftz.f32 	%f107, %f110;
$LDWendi___log2f_217_31:
	.loc	20	522	0
	mov.f32 	%f9, %f87;
	mov.f32 	%f10, %f97;
	mov.f32 	%f11, %f107;
$Lt_39_259842:
$Lt_39_88834:
	.loc	20	525	0
	mov.u32 	%r137, 0;
	setp.eq.s32 	%p81, %r38, %r137;
	@%p81 bra 	$Lt_39_260354;
	mov.u32 	%r138, 0;
	setp.ne.s32 	%p82, %r43, %r138;
	@%p82 bra 	$Lt_39_260354;
	.loc	20	57	0
	mov.u32 	%r139, 0;
	setp.ne.s32 	%p83, %r33, %r139;
	@%p83 bra 	$Lt_39_173570;
	.loc	20	59	0
	mov.f32 	%f111, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_29;
$Lt_39_173570:
	.loc	20	61	0
	mov.u32 	%r140, 64;
	setp.ne.s32 	%p84, %r33, %r140;
	@%p84 bra 	$Lt_39_173826;
	.loc	20	63	0
	mov.f32 	%f111, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_29;
$Lt_39_173826:
	.loc	20	65	0
	mov.u32 	%r141, 128;
	setp.ne.s32 	%p85, %r33, %r141;
	@%p85 bra 	$Lt_39_174082;
	.loc	20	68	0
	mov.f32 	%f111, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_29;
$Lt_39_174082:
	.loc	20	70	0
	mov.u32 	%r142, 192;
	setp.ne.s32 	%p86, %r33, %r142;
	@%p86 bra 	$Lt_39_174338;
	.loc	20	72	0
	mov.f32 	%f111, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_29;
$Lt_39_174338:
	.loc	20	76	0
	mov.f32 	%f111, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_217_29:
	.loc	20	107	0
	and.b32 	%r143, %r12, 2048;
	mov.s32 	%r144, 0;
	setp.ne.s32 	%p87, %r143, %r144;
	@!%p87 bra 	$Lt_39_243970;
	.loc	20	100	0
	ld.const.f32 	%f112, [kYCbCrOffset+0];
	bra.uni 	$Lt_39_243714;
$Lt_39_243970:
	ld.const.f32 	%f112, [kYCbCrFullRangeOffset+0];
$Lt_39_243714:
	.loc	20	107	0
	@!%p87 bra 	$Lt_39_244482;
	.loc	20	100	0
	ld.const.f32 	%f113, [kYCbCrOffset+4];
	bra.uni 	$Lt_39_244226;
$Lt_39_244482:
	ld.const.f32 	%f113, [kYCbCrFullRangeOffset+4];
$Lt_39_244226:
	.loc	20	107	0
	@!%p87 bra 	$Lt_39_244994;
	.loc	20	100	0
	ld.const.f32 	%f114, [kYCbCrOffset+8];
	bra.uni 	$Lt_39_244738;
$Lt_39_244994:
	ld.const.f32 	%f114, [kYCbCrFullRangeOffset+8];
$Lt_39_244738:
	.loc	20	527	0
	mov.f32 	%f115, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f116, %f111, %f115;
	fma.rn.ftz.f32 	%f9, %f116, %f112, %f9;
	fma.rn.ftz.f32 	%f10, %f116, %f113, %f10;
	fma.rn.ftz.f32 	%f11, %f116, %f114, %f11;
$Lt_39_260354:
$Lt_39_95234:
	.loc	20	525	0
	and.b32 	%r145, %r12, 12;
	and.b32 	%r146, %r25, 12;
	setp.eq.s32 	%p88, %r145, %r146;
	@%p88 bra 	$Lt_39_245762;
	.loc	20	532	0
	mov.u32 	%r147, 8;
	setp.ne.s32 	%p89, %r146, %r147;
	@%p89 bra 	$L_39_228610;
	mov.u32 	%r148, 12;
	setp.eq.s32 	%p90, %r145, %r148;
	@%p90 bra 	$Lt_39_261122;
$L_39_228610:
	mov.u32 	%r149, 12;
	setp.eq.s32 	%p91, %r146, %r149;
	@%p91 bra 	$Lt_39_261122;
	mov.u32 	%r150, 0;
	setp.ne.s32 	%p92, %r146, %r150;
	@%p92 bra 	$L_39_227842;
$Lt_39_261122:
$L_39_228098:
	.loc	20	57	0
	mov.u32 	%r151, 0;
	setp.ne.s32 	%p93, %r33, %r151;
	@%p93 bra 	$Lt_39_176898;
	.loc	20	59	0
	mov.f32 	%f117, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_25;
$Lt_39_176898:
	.loc	20	61	0
	mov.u32 	%r152, 64;
	setp.ne.s32 	%p94, %r33, %r152;
	@%p94 bra 	$Lt_39_177154;
	.loc	20	63	0
	mov.f32 	%f117, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_25;
$Lt_39_177154:
	.loc	20	65	0
	mov.u32 	%r153, 128;
	setp.ne.s32 	%p95, %r33, %r153;
	@%p95 bra 	$Lt_39_177410;
	.loc	20	68	0
	mov.f32 	%f117, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_25;
$Lt_39_177410:
	.loc	20	70	0
	mov.u32 	%r154, 192;
	setp.ne.s32 	%p96, %r33, %r154;
	@%p96 bra 	$Lt_39_177666;
	.loc	20	72	0
	mov.f32 	%f117, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_25;
$Lt_39_177666:
	.loc	20	76	0
	mov.f32 	%f117, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_217_25:
	.loc	20	536	0
	mov.f32 	%f7, %f117;
	bra.uni 	$Lt_39_245762;
$L_39_227842:
	.loc	20	540	0
	mov.s32 	%r155, 12;
	setp.eq.s32 	%p97, %r145, %r155;
	mov.s32 	%r156, 4;
	set.eq.u32.s32 	%r157, %r146, %r156;
	neg.s32 	%r158, %r157;
	selp.s32 	%r159, 1, 0, %p97;
	mov.s32 	%r160, 8;
	set.eq.u32.s32 	%r161, %r145, %r160;
	neg.s32 	%r162, %r161;
	or.b32 	%r163, %r159, %r162;
	and.b32 	%r164, %r158, %r163;
	mov.u32 	%r165, 0;
	setp.eq.s32 	%p98, %r164, %r165;
	@%p98 bra 	$Lt_39_246018;
	.loc	20	410	0
	mov.f32 	%f118, %f9;
	mov.f32 	%f119, %f118;
	mov.f32 	%f120, %f10;
	mov.f32 	%f121, %f120;
	mov.f32 	%f122, %f11;
	mov.f32 	%f123, %f122;
	.loc	20	57	0
	mov.s32 	%r166, 0;
	setp.eq.s32 	%p40, %r33, %r166;
	@!%p40 bra 	$Lt_39_178178;
	.loc	20	59	0
	mov.f32 	%f124, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_23;
$Lt_39_178178:
	.loc	20	61	0
	mov.u32 	%r167, 64;
	setp.ne.s32 	%p99, %r33, %r167;
	@%p99 bra 	$Lt_39_178434;
	.loc	20	63	0
	mov.f32 	%f124, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_23;
$Lt_39_178434:
	.loc	20	65	0
	mov.u32 	%r168, 128;
	setp.ne.s32 	%p100, %r33, %r168;
	@%p100 bra 	$Lt_39_178690;
	.loc	20	68	0
	mov.f32 	%f124, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_23;
$Lt_39_178690:
	.loc	20	70	0
	mov.u32 	%r169, 192;
	setp.ne.s32 	%p101, %r33, %r169;
	@%p101 bra 	$Lt_39_178946;
	.loc	20	72	0
	mov.f32 	%f124, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_23;
$Lt_39_178946:
	.loc	20	76	0
	mov.f32 	%f124, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_217_23:
	.loc	20	413	0
	mov.u32 	%r170, 0;
	setp.eq.s32 	%p102, %r38, %r170;
	@%p102 bra 	$Lt_39_246274;
	.loc	20	57	0
	@!%p40 bra 	$Lt_39_179458;
	.loc	20	59	0
	mov.f32 	%f125, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_21;
$Lt_39_179458:
	.loc	20	61	0
	mov.u32 	%r171, 64;
	setp.ne.s32 	%p103, %r33, %r171;
	@%p103 bra 	$Lt_39_179714;
	.loc	20	63	0
	mov.f32 	%f125, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_21;
$Lt_39_179714:
	.loc	20	65	0
	mov.u32 	%r172, 128;
	setp.ne.s32 	%p104, %r33, %r172;
	@%p104 bra 	$Lt_39_179970;
	.loc	20	68	0
	mov.f32 	%f125, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_21;
$Lt_39_179970:
	.loc	20	70	0
	mov.u32 	%r173, 192;
	setp.ne.s32 	%p105, %r33, %r173;
	@%p105 bra 	$Lt_39_180226;
	.loc	20	72	0
	mov.f32 	%f125, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_21;
$Lt_39_180226:
	.loc	20	76	0
	mov.f32 	%f125, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_217_21:
	.loc	20	118	0
	and.b32 	%r143, %r12, 2048;
	mov.s32 	%r174, 0;
	setp.ne.s32 	%p87, %r143, %r174;
	@!%p87 bra 	$Lt_39_247042;
	.loc	20	100	0
	ld.const.f32 	%f126, [kYCbCrOffset+0];
	bra.uni 	$Lt_39_246786;
$Lt_39_247042:
	ld.const.f32 	%f126, [kYCbCrFullRangeOffset+0];
$Lt_39_246786:
	.loc	20	118	0
	@!%p87 bra 	$Lt_39_247554;
	.loc	20	100	0
	ld.const.f32 	%f127, [kYCbCrOffset+4];
	bra.uni 	$Lt_39_247298;
$Lt_39_247554:
	ld.const.f32 	%f127, [kYCbCrFullRangeOffset+4];
$Lt_39_247298:
	.loc	20	118	0
	@!%p87 bra 	$Lt_39_248066;
	.loc	20	100	0
	ld.const.f32 	%f128, [kYCbCrOffset+8];
	bra.uni 	$Lt_39_247810;
$Lt_39_248066:
	ld.const.f32 	%f128, [kYCbCrFullRangeOffset+8];
$Lt_39_247810:
	.loc	20	415	0
	mov.f32 	%f129, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f130, %f125, %f129;
	mul.ftz.f32 	%f131, %f130, %f126;
	sub.ftz.f32 	%f119, %f118, %f131;
	mul.ftz.f32 	%f132, %f130, %f127;
	sub.ftz.f32 	%f121, %f120, %f132;
	mul.ftz.f32 	%f133, %f130, %f128;
	sub.ftz.f32 	%f123, %f122, %f133;
$Lt_39_246274:
	.loc	20	418	0
	rcp.approx.ftz.f32 	%f134, %f124;
	mul.ftz.f32 	%f135, %f134, %f7;
	mul.ftz.f32 	%f119, %f135, %f119;
	.loc	20	419	0
	mul.ftz.f32 	%f121, %f135, %f121;
	.loc	20	420	0
	mul.ftz.f32 	%f123, %f135, %f123;
	.loc	20	422	0
	mov.u32 	%r175, 0;
	setp.eq.s32 	%p106, %r38, %r175;
	@%p106 bra 	$Lt_39_248322;
	.loc	20	57	0
	@!%p40 bra 	$Lt_39_182530;
	.loc	20	59	0
	mov.f32 	%f136, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_17;
$Lt_39_182530:
	.loc	20	61	0
	mov.u32 	%r176, 64;
	setp.ne.s32 	%p107, %r33, %r176;
	@%p107 bra 	$Lt_39_182786;
	.loc	20	63	0
	mov.f32 	%f136, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_17;
$Lt_39_182786:
	.loc	20	65	0
	mov.u32 	%r177, 128;
	setp.ne.s32 	%p108, %r33, %r177;
	@%p108 bra 	$Lt_39_183042;
	.loc	20	68	0
	mov.f32 	%f136, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_17;
$Lt_39_183042:
	.loc	20	70	0
	mov.u32 	%r178, 192;
	setp.ne.s32 	%p109, %r33, %r178;
	@%p109 bra 	$Lt_39_183298;
	.loc	20	72	0
	mov.f32 	%f136, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_17;
$Lt_39_183298:
	.loc	20	76	0
	mov.f32 	%f136, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_217_17:
	.loc	20	107	0
	and.b32 	%r143, %r12, 2048;
	mov.s32 	%r179, 0;
	setp.ne.s32 	%p87, %r143, %r179;
	@!%p87 bra 	$Lt_39_249090;
	.loc	20	100	0
	ld.const.f32 	%f137, [kYCbCrOffset+0];
	bra.uni 	$Lt_39_248834;
$Lt_39_249090:
	ld.const.f32 	%f137, [kYCbCrFullRangeOffset+0];
$Lt_39_248834:
	.loc	20	107	0
	@!%p87 bra 	$Lt_39_249602;
	.loc	20	100	0
	ld.const.f32 	%f138, [kYCbCrOffset+4];
	bra.uni 	$Lt_39_249346;
$Lt_39_249602:
	ld.const.f32 	%f138, [kYCbCrFullRangeOffset+4];
$Lt_39_249346:
	.loc	20	107	0
	@!%p87 bra 	$Lt_39_250114;
	.loc	20	100	0
	ld.const.f32 	%f139, [kYCbCrOffset+8];
	bra.uni 	$Lt_39_249858;
$Lt_39_250114:
	ld.const.f32 	%f139, [kYCbCrFullRangeOffset+8];
$Lt_39_249858:
	.loc	20	424	0
	mov.f32 	%f140, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f141, %f136, %f140;
	fma.rn.ftz.f32 	%f119, %f141, %f137, %f119;
	fma.rn.ftz.f32 	%f121, %f141, %f138, %f121;
	fma.rn.ftz.f32 	%f123, %f141, %f139, %f123;
$Lt_39_248322:
	.loc	20	543	0
	mov.f32 	%f9, %f119;
	mov.f32 	%f10, %f121;
	mov.f32 	%f11, %f123;
	@!%p97 bra 	$Lt_39_245762;
	.loc	20	57	0
	@!%p40 bra 	$Lt_39_185602;
	.loc	20	59	0
	mov.f32 	%f117, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_13;
$Lt_39_185602:
	.loc	20	61	0
	mov.u32 	%r180, 64;
	setp.ne.s32 	%p110, %r33, %r180;
	@%p110 bra 	$Lt_39_185858;
	.loc	20	63	0
	mov.f32 	%f117, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_13;
$Lt_39_185858:
	.loc	20	65	0
	mov.u32 	%r181, 128;
	setp.ne.s32 	%p111, %r33, %r181;
	@%p111 bra 	$Lt_39_186114;
	.loc	20	68	0
	mov.f32 	%f117, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_13;
$Lt_39_186114:
	.loc	20	70	0
	mov.u32 	%r182, 192;
	setp.ne.s32 	%p112, %r33, %r182;
	@%p112 bra 	$Lt_39_186370;
	.loc	20	72	0
	mov.f32 	%f117, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_13;
$Lt_39_186370:
	.loc	20	76	0
	mov.f32 	%f117, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_217_13:
	.loc	20	546	0
	mov.f32 	%f7, %f117;
	bra.uni 	$Lt_39_245762;
$Lt_39_246018:
	.loc	20	433	0
	mov.f32 	%f142, %f7;
	mov.f32 	%f143, %f9;
	mov.f32 	%f144, %f143;
	mov.f32 	%f145, %f10;
	mov.f32 	%f146, %f145;
	mov.f32 	%f147, %f11;
	mov.f32 	%f148, %f147;
	.loc	20	435	0
	mov.u32 	%r183, 0;
	setp.eq.s32 	%p113, %r38, %r183;
	@%p113 bra 	$Lt_39_250882;
	.loc	20	57	0
	mov.u32 	%r184, 0;
	setp.ne.s32 	%p114, %r33, %r184;
	@%p114 bra 	$Lt_39_186882;
	.loc	20	59	0
	mov.f32 	%f149, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_11;
$Lt_39_186882:
	.loc	20	61	0
	mov.u32 	%r185, 64;
	setp.ne.s32 	%p115, %r33, %r185;
	@%p115 bra 	$Lt_39_187138;
	.loc	20	63	0
	mov.f32 	%f149, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_11;
$Lt_39_187138:
	.loc	20	65	0
	mov.u32 	%r186, 128;
	setp.ne.s32 	%p116, %r33, %r186;
	@%p116 bra 	$Lt_39_187394;
	.loc	20	68	0
	mov.f32 	%f149, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_11;
$Lt_39_187394:
	.loc	20	70	0
	mov.u32 	%r187, 192;
	setp.ne.s32 	%p117, %r33, %r187;
	@%p117 bra 	$Lt_39_187650;
	.loc	20	72	0
	mov.f32 	%f149, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_11;
$Lt_39_187650:
	.loc	20	76	0
	mov.f32 	%f149, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_217_11:
	.loc	20	118	0
	and.b32 	%r143, %r12, 2048;
	mov.s32 	%r188, 0;
	setp.ne.s32 	%p87, %r143, %r188;
	@!%p87 bra 	$Lt_39_251650;
	.loc	20	100	0
	ld.const.f32 	%f150, [kYCbCrOffset+0];
	bra.uni 	$Lt_39_251394;
$Lt_39_251650:
	ld.const.f32 	%f150, [kYCbCrFullRangeOffset+0];
$Lt_39_251394:
	.loc	20	118	0
	@!%p87 bra 	$Lt_39_252162;
	.loc	20	100	0
	ld.const.f32 	%f151, [kYCbCrOffset+4];
	bra.uni 	$Lt_39_251906;
$Lt_39_252162:
	ld.const.f32 	%f151, [kYCbCrFullRangeOffset+4];
$Lt_39_251906:
	.loc	20	118	0
	@!%p87 bra 	$Lt_39_252674;
	.loc	20	100	0
	ld.const.f32 	%f152, [kYCbCrOffset+8];
	bra.uni 	$Lt_39_252418;
$Lt_39_252674:
	ld.const.f32 	%f152, [kYCbCrFullRangeOffset+8];
$Lt_39_252418:
	.loc	20	437	0
	mov.f32 	%f153, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f154, %f149, %f153;
	mul.ftz.f32 	%f155, %f154, %f150;
	sub.ftz.f32 	%f144, %f143, %f155;
	mul.ftz.f32 	%f156, %f154, %f151;
	sub.ftz.f32 	%f146, %f145, %f156;
	mul.ftz.f32 	%f157, %f154, %f152;
	sub.ftz.f32 	%f148, %f147, %f157;
$Lt_39_250882:
	mov.f32 	%f158, 0fb70637bd;   	// -8e-006
	add.ftz.f32 	%f159, %f7, %f158;
	mov.f32 	%f160, 0f00000000;   	// 0
	setp.le.ftz.f32 	%p118, %f159, %f160;
	@!%p118 bra 	$Lt_39_253186;
	mov.f32 	%f148, 0f00000000;   	// 0
	mov.f32 	%f146, 0f00000000;   	// 0
	mov.f32 	%f144, 0f00000000;   	// 0
	mov.f32 	%f142, 0f00000000;   	// 0
	bra.uni 	$Lt_39_252930;
$Lt_39_253186:
	.loc	20	57	0
	mov.u32 	%r189, 0;
	setp.ne.s32 	%p119, %r33, %r189;
	@%p119 bra 	$Lt_39_189954;
	.loc	20	59	0
	mov.f32 	%f161, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_7;
$Lt_39_189954:
	.loc	20	61	0
	mov.u32 	%r190, 64;
	setp.ne.s32 	%p120, %r33, %r190;
	@%p120 bra 	$Lt_39_190210;
	.loc	20	63	0
	mov.f32 	%f161, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_7;
$Lt_39_190210:
	.loc	20	65	0
	mov.u32 	%r191, 128;
	setp.ne.s32 	%p121, %r33, %r191;
	@%p121 bra 	$Lt_39_190466;
	.loc	20	68	0
	mov.f32 	%f161, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_7;
$Lt_39_190466:
	.loc	20	70	0
	mov.u32 	%r192, 192;
	setp.ne.s32 	%p122, %r33, %r192;
	@%p122 bra 	$Lt_39_190722;
	.loc	20	72	0
	mov.f32 	%f161, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_7;
$Lt_39_190722:
	.loc	20	76	0
	mov.f32 	%f161, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_217_7:
	.loc	20	447	0
	div.approx.ftz.f32 	%f162, %f161, %f7;
	mul.ftz.f32 	%f144, %f162, %f144;
	.loc	20	448	0
	mul.ftz.f32 	%f146, %f162, %f146;
	.loc	20	449	0
	mul.ftz.f32 	%f148, %f162, %f148;
$Lt_39_252930:
	.loc	20	452	0
	mov.u32 	%r193, 0;
	setp.eq.s32 	%p123, %r38, %r193;
	@%p123 bra 	$Lt_39_253442;
	.loc	20	57	0
	mov.u32 	%r194, 0;
	setp.ne.s32 	%p124, %r33, %r194;
	@%p124 bra 	$Lt_39_191234;
	.loc	20	59	0
	mov.f32 	%f163, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_5;
$Lt_39_191234:
	.loc	20	61	0
	mov.u32 	%r195, 64;
	setp.ne.s32 	%p125, %r33, %r195;
	@%p125 bra 	$Lt_39_191490;
	.loc	20	63	0
	mov.f32 	%f163, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_5;
$Lt_39_191490:
	.loc	20	65	0
	mov.u32 	%r196, 128;
	setp.ne.s32 	%p126, %r33, %r196;
	@%p126 bra 	$Lt_39_191746;
	.loc	20	68	0
	mov.f32 	%f163, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_5;
$Lt_39_191746:
	.loc	20	70	0
	mov.u32 	%r197, 192;
	setp.ne.s32 	%p127, %r33, %r197;
	@%p127 bra 	$Lt_39_192002;
	.loc	20	72	0
	mov.f32 	%f163, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_5;
$Lt_39_192002:
	.loc	20	76	0
	mov.f32 	%f163, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_217_5:
	.loc	20	107	0
	and.b32 	%r143, %r12, 2048;
	mov.s32 	%r198, 0;
	setp.ne.s32 	%p87, %r143, %r198;
	@!%p87 bra 	$Lt_39_254210;
	.loc	20	100	0
	ld.const.f32 	%f164, [kYCbCrOffset+0];
	bra.uni 	$Lt_39_253954;
$Lt_39_254210:
	ld.const.f32 	%f164, [kYCbCrFullRangeOffset+0];
$Lt_39_253954:
	.loc	20	107	0
	@!%p87 bra 	$Lt_39_254722;
	.loc	20	100	0
	ld.const.f32 	%f165, [kYCbCrOffset+4];
	bra.uni 	$Lt_39_254466;
$Lt_39_254722:
	ld.const.f32 	%f165, [kYCbCrFullRangeOffset+4];
$Lt_39_254466:
	.loc	20	107	0
	@!%p87 bra 	$Lt_39_255234;
	.loc	20	100	0
	ld.const.f32 	%f166, [kYCbCrOffset+8];
	bra.uni 	$Lt_39_254978;
$Lt_39_255234:
	ld.const.f32 	%f166, [kYCbCrFullRangeOffset+8];
$Lt_39_254978:
	.loc	20	454	0
	mov.f32 	%f167, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f168, %f163, %f167;
	fma.rn.ftz.f32 	%f144, %f168, %f164, %f144;
	fma.rn.ftz.f32 	%f146, %f168, %f165, %f146;
	fma.rn.ftz.f32 	%f148, %f168, %f166, %f148;
$Lt_39_253442:
	.loc	20	551	0
	mov.f32 	%f7, %f142;
	mov.f32 	%f9, %f144;
	mov.f32 	%f10, %f146;
	mov.f32 	%f11, %f148;
$Lt_39_245762:
$L_39_227586:
$Lt_39_245250:
	.loc	20	540	0
	and.b32 	%r199, %r12, 4096;
	mov.u32 	%r200, 0;
	setp.ne.s32 	%p128, %r199, %r200;
	@%p128 bra 	$Lt_39_255490;
	.loc	21	268	0
	mov.f32 	%f169, %f9;
	.loc	21	269	0
	mov.f32 	%f170, %f7;
	.loc	20	558	0
	mov.f32 	%f7, %f11;
	mov.f32 	%f9, %f10;
	mov.f32 	%f10, %f169;
	mov.f32 	%f11, %f170;
$Lt_39_255490:
	@!%p8 bra 	$Lt_39_256002;
	.loc	20	57	0
	mov.u32 	%r201, 0;
	setp.ne.s32 	%p129, %r33, %r201;
	@%p129 bra 	$Lt_39_194562;
	.loc	20	59	0
	mov.f32 	%f171, 0f437f0000;   	// 255
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_1;
$Lt_39_194562:
	.loc	20	61	0
	mov.u32 	%r202, 64;
	setp.ne.s32 	%p130, %r33, %r202;
	@%p130 bra 	$Lt_39_194818;
	.loc	20	63	0
	mov.f32 	%f171, 0f447fc000;   	// 1023
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_1;
$Lt_39_194818:
	.loc	20	65	0
	mov.u32 	%r203, 128;
	setp.ne.s32 	%p131, %r33, %r203;
	@%p131 bra 	$Lt_39_195074;
	.loc	20	68	0
	mov.f32 	%f171, 0f47000000;   	// 32768
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_1;
$Lt_39_195074:
	.loc	20	70	0
	mov.u32 	%r204, 192;
	setp.ne.s32 	%p132, %r33, %r204;
	@%p132 bra 	$Lt_39_195330;
	.loc	20	72	0
	mov.f32 	%f171, 0fbf800000;   	// -1
	bra.uni 	$LDWendi__Z19MaxUnsignedBitValuei_217_1;
$Lt_39_195330:
	.loc	20	76	0
	mov.f32 	%f171, 0f3f800000;   	// 1
$LDWendi__Z19MaxUnsignedBitValuei_217_1:
	.loc	20	564	0
	mov.f32 	%f172, 0f3f000000;   	// 0.5
	add.ftz.f32 	%f173, %f7, %f172;
	mov.f32 	%f174, 0f00000000;   	// 0
	mov.f32 	%f175, 0f00000000;   	// 0
	setp.gt.ftz.f32 	%p133, %f173, %f175;
	selp.f32 	%f176, %f173, %f174, %p133;
	min.ftz.f32 	%f7, %f176, %f171;
	mov.f32 	%f177, 0f3f000000;   	// 0.5
	add.ftz.f32 	%f178, %f9, %f177;
	mov.f32 	%f179, 0f00000000;   	// 0
	mov.f32 	%f180, 0f00000000;   	// 0
	setp.gt.ftz.f32 	%p134, %f178, %f180;
	selp.f32 	%f181, %f178, %f179, %p134;
	min.ftz.f32 	%f9, %f181, %f171;
	mov.f32 	%f182, 0f3f000000;   	// 0.5
	add.ftz.f32 	%f183, %f10, %f182;
	mov.f32 	%f184, 0f00000000;   	// 0
	mov.f32 	%f185, 0f00000000;   	// 0
	setp.gt.ftz.f32 	%p135, %f183, %f185;
	selp.f32 	%f186, %f183, %f184, %p135;
	min.ftz.f32 	%f10, %f186, %f171;
	mov.f32 	%f187, 0f3f000000;   	// 0.5
	add.ftz.f32 	%f188, %f11, %f187;
	mov.f32 	%f189, 0f00000000;   	// 0
	mov.f32 	%f190, 0f00000000;   	// 0
	setp.gt.ftz.f32 	%p136, %f188, %f190;
	selp.f32 	%f191, %f188, %f189, %p136;
	min.ftz.f32 	%f11, %f191, %f171;
$Lt_39_256002:
	.loc	22	113	0
	mov.f32 	%f192, %f7;
	st.param.f32 	[__cudaretf__Z30ConvertPixel_422_8u_To_444_32f6uchar414IR_PixelFormatS0_+0], %f192;
	mov.f32 	%f193, %f9;
	st.param.f32 	[__cudaretf__Z30ConvertPixel_422_8u_To_444_32f6uchar414IR_PixelFormatS0_+4], %f193;
	mov.f32 	%f194, %f10;
	st.param.f32 	[__cudaretf__Z30ConvertPixel_422_8u_To_444_32f6uchar414IR_PixelFormatS0_+8], %f194;
	mov.f32 	%f195, %f11;
	st.param.f32 	[__cudaretf__Z30ConvertPixel_422_8u_To_444_32f6uchar414IR_PixelFormatS0_+12], %f195;
	ret;
$LDWend__Z30ConvertPixel_422_8u_To_444_32f6uchar414IR_PixelFormatS0_:
	} // _Z30ConvertPixel_422_8u_To_444_32f6uchar414IR_PixelFormatS0_

	.visible .func (.param .align 4 .b8 __cudaretf__Z6Read2DI6uchar4ET_PKS1_iii[4]) _Z6Read2DI6uchar4ET_PKS1_iii (.param .u64 __cudaparmf1__Z6Read2DI6uchar4ET_PKS1_iii, .param .s32 __cudaparmf2__Z6Read2DI6uchar4ET_PKS1_iii, .param .s32 __cudaparmf3__Z6Read2DI6uchar4ET_PKS1_iii, .param .s32 __cudaparmf4__Z6Read2DI6uchar4ET_PKS1_iii)
	{
	.reg .u32 %r<14>;
	.reg .u64 %rd<7>;
	.loc	19	114	0
$LDWbegin__Z6Read2DI6uchar4ET_PKS1_iii:
	ld.param.u64 	%rd1, [__cudaparmf1__Z6Read2DI6uchar4ET_PKS1_iii];
	mov.s64 	%rd2, %rd1;
	ld.param.u32 	%r1, [__cudaparmf2__Z6Read2DI6uchar4ET_PKS1_iii];
	mov.s32 	%r2, %r1;
	ld.param.u32 	%r3, [__cudaparmf3__Z6Read2DI6uchar4ET_PKS1_iii];
	mov.s32 	%r4, %r3;
	ld.param.u32 	%r5, [__cudaparmf4__Z6Read2DI6uchar4ET_PKS1_iii];
	mov.s32 	%r6, %r5;
	.loc	19	115	0
	mul.lo.s32 	%r7, %r2, %r6;
	add.s32 	%r8, %r4, %r7;
	cvt.s64.s32 	%rd3, %r8;
	mul.wide.s32 	%rd4, %r8, 4;
	add.u64 	%rd5, %rd2, %rd4;
	ld.v4.u8 	{%r9,%r10,%r11,%r12}, [%rd5+0];
	st.param.u8 	[__cudaretf__Z6Read2DI6uchar4ET_PKS1_iii+0], %r9;
	st.param.u8 	[__cudaretf__Z6Read2DI6uchar4ET_PKS1_iii+1], %r10;
	st.param.u8 	[__cudaretf__Z6Read2DI6uchar4ET_PKS1_iii+2], %r11;
	st.param.u8 	[__cudaretf__Z6Read2DI6uchar4ET_PKS1_iii+3], %r12;
	ret;
$LDWend__Z6Read2DI6uchar4ET_PKS1_iii:
	} // _Z6Read2DI6uchar4ET_PKS1_iii

	.visible .func (.param .align 16 .b8 __cudaretf__Z6Read2DI5uint4ET_PKS1_iii[16]) _Z6Read2DI5uint4ET_PKS1_iii (.param .u64 __cudaparmf1__Z6Read2DI5uint4ET_PKS1_iii, .param .s32 __cudaparmf2__Z6Read2DI5uint4ET_PKS1_iii, .param .s32 __cudaparmf3__Z6Read2DI5uint4ET_PKS1_iii, .param .s32 __cudaparmf4__Z6Read2DI5uint4ET_PKS1_iii)
	{
	.reg .u32 %r<14>;
	.reg .u64 %rd<7>;
	.loc	19	114	0
$LDWbegin__Z6Read2DI5uint4ET_PKS1_iii:
	ld.param.u64 	%rd1, [__cudaparmf1__Z6Read2DI5uint4ET_PKS1_iii];
	mov.s64 	%rd2, %rd1;
	ld.param.u32 	%r1, [__cudaparmf2__Z6Read2DI5uint4ET_PKS1_iii];
	mov.s32 	%r2, %r1;
	ld.param.u32 	%r3, [__cudaparmf3__Z6Read2DI5uint4ET_PKS1_iii];
	mov.s32 	%r4, %r3;
	ld.param.u32 	%r5, [__cudaparmf4__Z6Read2DI5uint4ET_PKS1_iii];
	mov.s32 	%r6, %r5;
	.loc	19	115	0
	mul.lo.s32 	%r7, %r2, %r6;
	add.s32 	%r8, %r4, %r7;
	cvt.s64.s32 	%rd3, %r8;
	mul.wide.s32 	%rd4, %r8, 16;
	add.u64 	%rd5, %rd2, %rd4;
	ld.v4.u32 	{%r9,%r10,%r11,%r12}, [%rd5+0];
	st.param.u32 	[__cudaretf__Z6Read2DI5uint4ET_PKS1_iii+0], %r9;
	st.param.u32 	[__cudaretf__Z6Read2DI5uint4ET_PKS1_iii+4], %r10;
	st.param.u32 	[__cudaretf__Z6Read2DI5uint4ET_PKS1_iii+8], %r11;
	st.param.u32 	[__cudaretf__Z6Read2DI5uint4ET_PKS1_iii+12], %r12;
	ret;
$LDWend__Z6Read2DI5uint4ET_PKS1_iii:
	} // _Z6Read2DI5uint4ET_PKS1_iii

	.visible .func (.param .u64 __cudaretf__ZpLI6float4ERT_S2_f) _ZpLI6float4ERT_S2_f (.param .u64 __cudaparmf1__ZpLI6float4ERT_S2_f, .param .f32 __cudaparmf2__ZpLI6float4ERT_S2_f)
	{
	.reg .u64 %rd<5>;
	.reg .f32 %f<12>;
	.loc	21	91	0
$LDWbegin__ZpLI6float4ERT_S2_f:
	ld.param.u64 	%rd1, [__cudaparmf1__ZpLI6float4ERT_S2_f];
	mov.s64 	%rd2, %rd1;
	ld.param.f32 	%f1, [__cudaparmf2__ZpLI6float4ERT_S2_f];
	mov.f32 	%f2, %f1;
	ld.v4.f32 	{%f3,%f4,%f5,%f6}, [%rd2+0];
	.loc	21	92	0
	add.ftz.f32 	%f7, %f3, %f2;
	.loc	21	93	0
	add.ftz.f32 	%f8, %f4, %f2;
	.loc	21	94	0
	add.ftz.f32 	%f9, %f5, %f2;
	.loc	21	95	0
	add.ftz.f32 	%f10, %f6, %f2;
	st.v4.f32 	[%rd2+0], {%f7,%f8,%f9,%f10};
	.loc	21	96	0
	mov.s64 	%rd3, %rd2;
	st.param.u64 	[__cudaretf__ZpLI6float4ERT_S2_f], %rd3;
	ret;
$LDWend__ZpLI6float4ERT_S2_f:
	} // _ZpLI6float4ERT_S2_f

	.entry PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel (
		.param .u64 __cudaparm_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inSrc,
		.param .s32 __cudaparm_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inSrcPitch,
		.param .u32 __cudaparm_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel___T289,
		.param .u64 __cudaparm_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDest,
		.param .s32 __cudaparm_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDestPitch,
		.param .u32 __cudaparm_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDestDeviceFormat,
		.param .s32 __cudaparm_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inWidth,
		.param .s32 __cudaparm_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inHeight)
	{
	.reg .u32 %r<82>;
	.reg .u64 %rd<32>;
	.reg .f32 %f<170>;
	.reg .pred %p<9>;
	.loc	22	199	0
$LDWbegin_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel:
	cvt.s32.u32 	%r1, %ctaid.y;
	cvt.s32.u32 	%r2, %ntid.y;
	mul.lo.s32 	%r3, %r1, %r2;
	mov.u32 	%r4, %tid.y;
	add.u32 	%r5, %r3, %r4;
	ld.param.s32 	%r6, [__cudaparm_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inHeight];
	setp.le.s32 	%p1, %r6, %r5;
	@%p1 bra 	$Lt_43_1476866;
	cvt.s32.u32 	%r7, %ctaid.x;
	cvt.s32.u32 	%r8, %ntid.x;
	mul.lo.s32 	%r9, %r7, %r8;
	mov.u32 	%r10, %tid.x;
	add.u32 	%r11, %r9, %r10;
	shl.b32 	%r12, %r11, 1;
	ld.param.s32 	%r13, [__cudaparm_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inWidth];
	add.s32 	%r14, %r12, 2;
	setp.le.s32 	%p2, %r13, %r14;
	@%p2 bra 	$Lt_43_1435138;
	.loc	19	115	0
	ld.param.u64 	%rd1, [__cudaparm_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inSrc];
	ld.param.s32 	%r15, [__cudaparm_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inSrcPitch];
	mul.lo.s32 	%r16, %r15, %r5;
	add.s32 	%r17, %r11, %r16;
	cvt.s64.s32 	%rd2, %r17;
	mul.wide.s32 	%rd3, %r17, 4;
	add.u64 	%rd4, %rd1, %rd3;
	ld.global.v4.u8 	{%r18,%r19,%r20,%r21}, [%rd4+0];
	ld.global.u8 	%r22, [%rd4+4];
	ld.global.u8 	%r23, [%rd4+6];
	.loc	20	469	0
	cvt.rn.f32.u32 	%f1, %r20;
	cvt.rn.f32.u32 	%f2, %r18;
	cvt.rn.f32.u32 	%f3, %r19;
	.loc	22	60	0
	mov.f32 	%f4, 0f437f0000;     	// 255
	mov.f32 	%f5, 0f437f0000;     	// 255
	div.approx.ftz.f32 	%f6, %f4, %f5;
	ld.const.f32 	%f7, [kYCbCrOffset+4];
	mul.ftz.f32 	%f8, %f6, %f7;
	sub.ftz.f32 	%f9, %f2, %f8;
	ld.const.f32 	%f10, [kYCbCrOffset+0];
	mul.ftz.f32 	%f11, %f6, %f10;
	sub.ftz.f32 	%f12, %f3, %f11;
	ld.const.f32 	%f13, [kYCbCrOffset+8];
	mul.ftz.f32 	%f14, %f6, %f13;
	sub.ftz.f32 	%f15, %f1, %f14;
	ld.global.f32 	%f16, [k601YCbCr_To_RGB32f+32];
	ld.global.f32 	%f17, [k601YCbCr_To_RGB32f+24];
	ld.global.f32 	%f18, [k601YCbCr_To_RGB32f+28];
	mul.ftz.f32 	%f19, %f9, %f18;
	fma.rn.ftz.f32 	%f20, %f17, %f12, %f19;
	fma.rn.ftz.f32 	%f21, %f16, %f15, %f20;
	ld.global.f32 	%f22, [k601YCbCr_To_RGB32f+20];
	ld.global.f32 	%f23, [k601YCbCr_To_RGB32f+12];
	ld.global.f32 	%f24, [k601YCbCr_To_RGB32f+16];
	mul.ftz.f32 	%f25, %f9, %f24;
	fma.rn.ftz.f32 	%f26, %f23, %f12, %f25;
	fma.rn.ftz.f32 	%f27, %f22, %f15, %f26;
	ld.global.f32 	%f28, [k601YCbCr_To_RGB32f+8];
	ld.global.f32 	%f29, [k601YCbCr_To_RGB32f+0];
	ld.global.f32 	%f30, [k601YCbCr_To_RGB32f+4];
	mul.ftz.f32 	%f31, %f9, %f30;
	fma.rn.ftz.f32 	%f32, %f29, %f12, %f31;
	fma.rn.ftz.f32 	%f33, %f28, %f15, %f32;
	.loc	20	469	0
	cvt.rn.f32.u32 	%f34, %r23;
	add.ftz.f32 	%f35, %f34, %f1;
	cvt.rzi.ftz.u32.f32 	%r24, %f35;
	shr.u32 	%r25, %r24, 1;
	cvt.rn.f32.u32 	%f36, %r25;
	cvt.rn.f32.u32 	%f37, %r22;
	add.ftz.f32 	%f38, %f37, %f2;
	cvt.rzi.ftz.u32.f32 	%r26, %f38;
	shr.u32 	%r27, %r26, 1;
	cvt.rn.f32.u32 	%f39, %r27;
	cvt.rn.f32.u32 	%f40, %r21;
	.loc	20	490	0
	mul.ftz.f32 	%f41, %f6, %f7;
	sub.ftz.f32 	%f42, %f39, %f41;
	mul.ftz.f32 	%f43, %f6, %f10;
	sub.ftz.f32 	%f44, %f40, %f43;
	mul.ftz.f32 	%f45, %f6, %f13;
	sub.ftz.f32 	%f46, %f36, %f45;
	mul.ftz.f32 	%f47, %f42, %f24;
	fma.rn.ftz.f32 	%f48, %f23, %f44, %f47;
	fma.rn.ftz.f32 	%f49, %f22, %f46, %f48;
	mul.ftz.f32 	%f50, %f42, %f18;
	fma.rn.ftz.f32 	%f51, %f17, %f44, %f50;
	fma.rn.ftz.f32 	%f52, %f16, %f46, %f51;
	.loc	21	268	0
	mul.ftz.f32 	%f53, %f42, %f30;
	fma.rn.ftz.f32 	%f54, %f29, %f44, %f53;
	fma.rn.ftz.f32 	%f55, %f28, %f46, %f54;
	.loc	22	199	0
	ld.param.s32 	%r28, [__cudaparm_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDestDeviceFormat];
	mov.s32 	%r29, 0;
	setp.eq.s32 	%p3, %r28, %r29;
	ld.param.s32 	%r30, [__cudaparm_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDestPitch];
	mul.lo.s32 	%r31, %r30, %r5;
	add.s32 	%r32, %r12, %r31;
	cvt.s64.s32 	%rd5, %r32;
	ld.param.u64 	%rd6, [__cudaparm_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDest];
	@!%p3 bra 	$Lt_43_1451010;
	.loc	19	126	0
	mul.lo.u64 	%rd7, %rd5, 8;
	add.u64 	%rd8, %rd6, %rd7;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f21;
	mov.b32		%r33, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f27;
	mov.b32		%r34, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f33;
	mov.b32		%r35, %b1; }
	mov.f32 	%f56, 0f3f800000;    	// 1
	mov.f32 	%f57, 0f437f0000;    	// 255
	div.approx.ftz.f32 	%f58, %f56, %f57;
	mov.f32 	%f59, 0f437f0000;    	// 255
	mul.ftz.f32 	%f60, %f58, %f59;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f60;
	mov.b32		%r36, %b1; }
	st.global.v4.u16 	[%rd8+0], {%r33,%r34,%r35,%r36};
	.loc	22	199	0
	bra.uni 	$Lt_43_1450754;
$Lt_43_1451010:
	.loc	19	126	0
	mul.lo.u64 	%rd9, %rd5, 16;
	add.u64 	%rd10, %rd6, %rd9;
	mov.f32 	%f61, 0f3f800000;    	// 1
	mov.f32 	%f62, 0f437f0000;    	// 255
	div.approx.ftz.f32 	%f63, %f61, %f62;
	mov.f32 	%f64, 0f437f0000;    	// 255
	mul.ftz.f32 	%f60, %f63, %f64;
	st.global.v4.f32 	[%rd10+0], {%f21,%f27,%f33,%f60};
$Lt_43_1450754:
	.loc	22	199	0
	@!%p3 bra 	$Lt_43_1451522;
	.loc	19	126	0
	mul.lo.u64 	%rd11, %rd5, 8;
	add.u64 	%rd8, %rd6, %rd11;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f52;
	mov.b32		%r37, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f49;
	mov.b32		%r38, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f55;
	mov.b32		%r39, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f60;
	mov.b32		%r40, %b1; }
	st.global.v4.u16 	[%rd8+8], {%r37,%r38,%r39,%r40};
	.loc	22	199	0
	bra.uni 	$Lt_43_1476866;
$Lt_43_1451522:
	.loc	19	126	0
	mul.lo.u64 	%rd12, %rd5, 16;
	add.u64 	%rd10, %rd6, %rd12;
	st.global.v4.f32 	[%rd10+16], {%f52,%f49,%f55,%f60};
	bra.uni 	$Lt_43_1476866;
$Lt_43_1435138:
	.loc	22	199	0
	add.s32 	%r41, %r12, 1;
	setp.le.s32 	%p4, %r13, %r41;
	@%p4 bra 	$Lt_43_1452034;
	.loc	19	115	0
	ld.param.u64 	%rd13, [__cudaparm_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inSrc];
	ld.param.s32 	%r42, [__cudaparm_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inSrcPitch];
	mul.lo.s32 	%r43, %r42, %r5;
	add.s32 	%r44, %r11, %r43;
	cvt.s64.s32 	%rd14, %r44;
	mul.wide.s32 	%rd15, %r44, 4;
	add.u64 	%rd16, %rd13, %rd15;
	ld.global.v4.u8 	{%r45,%r46,%r47,%r48}, [%rd16+0];
	.loc	20	469	0
	cvt.rn.f32.u32 	%f65, %r47;
	cvt.rn.f32.u32 	%f66, %r45;
	cvt.rn.f32.u32 	%f67, %r46;
	.loc	22	60	0
	mov.f32 	%f68, 0f437f0000;    	// 255
	mov.f32 	%f69, 0f437f0000;    	// 255
	div.approx.ftz.f32 	%f70, %f68, %f69;
	ld.const.f32 	%f71, [kYCbCrOffset+4];
	mul.ftz.f32 	%f72, %f70, %f71;
	sub.ftz.f32 	%f73, %f66, %f72;
	ld.const.f32 	%f74, [kYCbCrOffset+0];
	mul.ftz.f32 	%f75, %f70, %f74;
	sub.ftz.f32 	%f76, %f67, %f75;
	ld.const.f32 	%f77, [kYCbCrOffset+8];
	mul.ftz.f32 	%f78, %f70, %f77;
	sub.ftz.f32 	%f79, %f65, %f78;
	ld.global.f32 	%f80, [k601YCbCr_To_RGB32f+32];
	ld.global.f32 	%f81, [k601YCbCr_To_RGB32f+24];
	ld.global.f32 	%f82, [k601YCbCr_To_RGB32f+28];
	mul.ftz.f32 	%f83, %f73, %f82;
	fma.rn.ftz.f32 	%f84, %f81, %f76, %f83;
	fma.rn.ftz.f32 	%f85, %f80, %f79, %f84;
	ld.global.f32 	%f86, [k601YCbCr_To_RGB32f+20];
	ld.global.f32 	%f87, [k601YCbCr_To_RGB32f+12];
	ld.global.f32 	%f88, [k601YCbCr_To_RGB32f+16];
	mul.ftz.f32 	%f89, %f73, %f88;
	fma.rn.ftz.f32 	%f90, %f87, %f76, %f89;
	fma.rn.ftz.f32 	%f91, %f86, %f79, %f90;
	ld.global.f32 	%f92, [k601YCbCr_To_RGB32f+8];
	ld.global.f32 	%f93, [k601YCbCr_To_RGB32f+0];
	ld.global.f32 	%f94, [k601YCbCr_To_RGB32f+4];
	mul.ftz.f32 	%f95, %f73, %f94;
	fma.rn.ftz.f32 	%f96, %f93, %f76, %f95;
	fma.rn.ftz.f32 	%f97, %f92, %f79, %f96;
	.loc	20	469	0
	add.ftz.f32 	%f98, %f65, %f65;
	cvt.rzi.ftz.u32.f32 	%r49, %f98;
	shr.u32 	%r50, %r49, 1;
	cvt.rn.f32.u32 	%f99, %r50;
	add.ftz.f32 	%f100, %f66, %f66;
	cvt.rzi.ftz.u32.f32 	%r51, %f100;
	shr.u32 	%r52, %r51, 1;
	cvt.rn.f32.u32 	%f101, %r52;
	cvt.rn.f32.u32 	%f102, %r48;
	.loc	20	490	0
	mul.ftz.f32 	%f103, %f70, %f71;
	sub.ftz.f32 	%f104, %f101, %f103;
	mul.ftz.f32 	%f105, %f70, %f74;
	sub.ftz.f32 	%f106, %f102, %f105;
	mul.ftz.f32 	%f107, %f70, %f77;
	sub.ftz.f32 	%f108, %f99, %f107;
	mul.ftz.f32 	%f109, %f104, %f88;
	fma.rn.ftz.f32 	%f110, %f87, %f106, %f109;
	fma.rn.ftz.f32 	%f111, %f86, %f108, %f110;
	mul.ftz.f32 	%f112, %f104, %f82;
	fma.rn.ftz.f32 	%f113, %f81, %f106, %f112;
	fma.rn.ftz.f32 	%f114, %f80, %f108, %f113;
	.loc	21	268	0
	mul.ftz.f32 	%f115, %f104, %f94;
	fma.rn.ftz.f32 	%f116, %f93, %f106, %f115;
	fma.rn.ftz.f32 	%f117, %f92, %f108, %f116;
	.loc	22	199	0
	ld.param.s32 	%r53, [__cudaparm_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDestDeviceFormat];
	mov.s32 	%r54, 0;
	setp.eq.s32 	%p5, %r53, %r54;
	ld.param.s32 	%r55, [__cudaparm_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDestPitch];
	mul.lo.s32 	%r56, %r55, %r5;
	add.s32 	%r57, %r12, %r56;
	cvt.s64.s32 	%rd17, %r57;
	ld.param.u64 	%rd18, [__cudaparm_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDest];
	@!%p5 bra 	$Lt_43_1467906;
	.loc	19	126	0
	mul.lo.u64 	%rd19, %rd17, 8;
	add.u64 	%rd8, %rd18, %rd19;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f85;
	mov.b32		%r58, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f91;
	mov.b32		%r59, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f97;
	mov.b32		%r60, %b1; }
	mov.f32 	%f118, 0f3f800000;   	// 1
	mov.f32 	%f119, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f120, %f118, %f119;
	mov.f32 	%f121, 0f437f0000;   	// 255
	mul.ftz.f32 	%f60, %f120, %f121;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f60;
	mov.b32		%r61, %b1; }
	st.global.v4.u16 	[%rd8+0], {%r58,%r59,%r60,%r61};
	.loc	22	199	0
	bra.uni 	$Lt_43_1467650;
$Lt_43_1467906:
	.loc	19	126	0
	mul.lo.u64 	%rd20, %rd17, 16;
	add.u64 	%rd10, %rd18, %rd20;
	mov.f32 	%f122, 0f3f800000;   	// 1
	mov.f32 	%f123, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f124, %f122, %f123;
	mov.f32 	%f125, 0f437f0000;   	// 255
	mul.ftz.f32 	%f60, %f124, %f125;
	st.global.v4.f32 	[%rd10+0], {%f85,%f91,%f97,%f60};
$Lt_43_1467650:
	.loc	22	199	0
	@!%p5 bra 	$Lt_43_1468418;
	.loc	19	126	0
	mul.lo.u64 	%rd21, %rd17, 8;
	add.u64 	%rd8, %rd18, %rd21;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f114;
	mov.b32		%r62, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f111;
	mov.b32		%r63, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f117;
	mov.b32		%r64, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f60;
	mov.b32		%r65, %b1; }
	st.global.v4.u16 	[%rd8+8], {%r62,%r63,%r64,%r65};
	.loc	22	199	0
	bra.uni 	$Lt_43_1476866;
$Lt_43_1468418:
	.loc	19	126	0
	mul.lo.u64 	%rd22, %rd17, 16;
	add.u64 	%rd10, %rd18, %rd22;
	st.global.v4.f32 	[%rd10+16], {%f114,%f111,%f117,%f60};
	bra.uni 	$Lt_43_1476866;
$Lt_43_1452034:
	.loc	22	199	0
	setp.ge.s32 	%p6, %r12, %r13;
	@%p6 bra 	$Lt_43_1476866;
	.loc	20	469	0
	ld.param.u64 	%rd23, [__cudaparm_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inSrc];
	ld.param.s32 	%r66, [__cudaparm_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inSrcPitch];
	mul.lo.s32 	%r67, %r66, %r5;
	add.s32 	%r68, %r11, %r67;
	cvt.s64.s32 	%rd24, %r68;
	mul.wide.s32 	%rd25, %r68, 4;
	add.u64 	%rd26, %rd23, %rd25;
	ld.global.v4.u8 	{%r69,%r70,%r71,_}, [%rd26+0];
	cvt.rn.f32.u32 	%f126, %r71;
	cvt.rn.f32.u32 	%f127, %r69;
	cvt.rn.f32.u32 	%f128, %r70;
	.loc	20	490	0
	mov.f32 	%f129, 0f437f0000;   	// 255
	mov.f32 	%f130, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f131, %f129, %f130;
	ld.const.f32 	%f132, [kYCbCrOffset+4];
	mul.ftz.f32 	%f133, %f131, %f132;
	sub.ftz.f32 	%f134, %f127, %f133;
	ld.const.f32 	%f135, [kYCbCrOffset+0];
	mul.ftz.f32 	%f136, %f131, %f135;
	sub.ftz.f32 	%f137, %f128, %f136;
	ld.const.f32 	%f138, [kYCbCrOffset+8];
	mul.ftz.f32 	%f139, %f131, %f138;
	sub.ftz.f32 	%f140, %f126, %f139;
	ld.global.f32 	%f141, [k601YCbCr_To_RGB32f+16];
	mul.ftz.f32 	%f142, %f141, %f134;
	ld.global.f32 	%f143, [k601YCbCr_To_RGB32f+12];
	fma.rn.ftz.f32 	%f144, %f143, %f137, %f142;
	ld.global.f32 	%f145, [k601YCbCr_To_RGB32f+20];
	fma.rn.ftz.f32 	%f146, %f145, %f140, %f144;
	.loc	21	268	0
	ld.global.f32 	%f147, [k601YCbCr_To_RGB32f+4];
	mul.ftz.f32 	%f148, %f147, %f134;
	ld.global.f32 	%f149, [k601YCbCr_To_RGB32f+0];
	fma.rn.ftz.f32 	%f150, %f149, %f137, %f148;
	ld.global.f32 	%f151, [k601YCbCr_To_RGB32f+8];
	fma.rn.ftz.f32 	%f152, %f151, %f140, %f150;
	.loc	22	199	0
	ld.param.s32 	%r72, [__cudaparm_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDestPitch];
	mul.lo.s32 	%r73, %r72, %r5;
	ld.global.f32 	%f153, [k601YCbCr_To_RGB32f+28];
	mul.ftz.f32 	%f154, %f153, %f134;
	add.s32 	%r74, %r12, %r73;
	ld.global.f32 	%f155, [k601YCbCr_To_RGB32f+24];
	fma.rn.ftz.f32 	%f156, %f155, %f137, %f154;
	cvt.s64.s32 	%rd27, %r74;
	ld.global.f32 	%f157, [k601YCbCr_To_RGB32f+32];
	fma.rn.ftz.f32 	%f158, %f157, %f140, %f156;
	ld.param.u64 	%rd28, [__cudaparm_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDest];
	ld.param.s32 	%r75, [__cudaparm_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDestDeviceFormat];
	mov.u32 	%r76, 0;
	setp.ne.s32 	%p7, %r75, %r76;
	@%p7 bra 	$Lt_43_1477122;
	.loc	19	126	0
	mul.lo.u64 	%rd29, %rd27, 8;
	add.u64 	%rd8, %rd28, %rd29;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f158;
	mov.b32		%r77, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f146;
	mov.b32		%r78, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f152;
	mov.b32		%r79, %b1; }
	mov.f32 	%f159, 0f3f800000;   	// 1
	mov.f32 	%f160, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f161, %f159, %f160;
	mov.f32 	%f162, 0f437f0000;   	// 255
	mul.ftz.f32 	%f163, %f161, %f162;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f163;
	mov.b32		%r80, %b1; }
	st.global.v4.u16 	[%rd8+0], {%r77,%r78,%r79,%r80};
	.loc	22	199	0
	bra.uni 	$Lt_43_1476866;
$Lt_43_1477122:
	.loc	19	126	0
	mul.lo.u64 	%rd30, %rd27, 16;
	add.u64 	%rd10, %rd28, %rd30;
	mov.f32 	%f164, 0f3f800000;   	// 1
	mov.f32 	%f165, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f166, %f164, %f165;
	mov.f32 	%f167, 0f437f0000;   	// 255
	mul.ftz.f32 	%f168, %f166, %f167;
	st.global.v4.f32 	[%rd10+0], {%f158,%f146,%f152,%f168};
$Lt_43_1476866:
$Lt_43_1468674:
$Lt_43_1451778:
$Lt_43_1434882:
$Lt_43_1434370:
	.loc	22	199	0
	exit;
$LDWend_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel:
	} // PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel

	.entry PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel (
		.param .u64 __cudaparm_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inSrc,
		.param .s32 __cudaparm_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inSrcPitch,
		.param .u32 __cudaparm_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel___T2101,
		.param .u64 __cudaparm_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDest,
		.param .s32 __cudaparm_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDestPitch,
		.param .u32 __cudaparm_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDestDeviceFormat,
		.param .s32 __cudaparm_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inWidth,
		.param .s32 __cudaparm_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inHeight)
	{
	.reg .u32 %r<82>;
	.reg .u64 %rd<32>;
	.reg .f32 %f<170>;
	.reg .pred %p<9>;
	.loc	22	200	0
$LDWbegin_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel:
	cvt.s32.u32 	%r1, %ctaid.y;
	cvt.s32.u32 	%r2, %ntid.y;
	mul.lo.s32 	%r3, %r1, %r2;
	mov.u32 	%r4, %tid.y;
	add.u32 	%r5, %r3, %r4;
	ld.param.s32 	%r6, [__cudaparm_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inHeight];
	setp.le.s32 	%p1, %r6, %r5;
	@%p1 bra 	$Lt_44_1476866;
	cvt.s32.u32 	%r7, %ctaid.x;
	cvt.s32.u32 	%r8, %ntid.x;
	mul.lo.s32 	%r9, %r7, %r8;
	mov.u32 	%r10, %tid.x;
	add.u32 	%r11, %r9, %r10;
	shl.b32 	%r12, %r11, 1;
	ld.param.s32 	%r13, [__cudaparm_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inWidth];
	add.s32 	%r14, %r12, 2;
	setp.le.s32 	%p2, %r13, %r14;
	@%p2 bra 	$Lt_44_1435138;
	.loc	19	115	0
	ld.param.u64 	%rd1, [__cudaparm_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inSrc];
	ld.param.s32 	%r15, [__cudaparm_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inSrcPitch];
	mul.lo.s32 	%r16, %r15, %r5;
	add.s32 	%r17, %r11, %r16;
	cvt.s64.s32 	%rd2, %r17;
	mul.wide.s32 	%rd3, %r17, 4;
	add.u64 	%rd4, %rd1, %rd3;
	ld.global.v4.u8 	{%r18,%r19,%r20,%r21}, [%rd4+0];
	ld.global.u8 	%r22, [%rd4+4];
	ld.global.u8 	%r23, [%rd4+6];
	.loc	20	469	0
	cvt.rn.f32.u32 	%f1, %r20;
	cvt.rn.f32.u32 	%f2, %r18;
	cvt.rn.f32.u32 	%f3, %r19;
	.loc	22	60	0
	mov.f32 	%f4, 0f437f0000;     	// 255
	mov.f32 	%f5, 0f437f0000;     	// 255
	div.approx.ftz.f32 	%f6, %f4, %f5;
	ld.const.f32 	%f7, [kYCbCrOffset+4];
	mul.ftz.f32 	%f8, %f6, %f7;
	sub.ftz.f32 	%f9, %f2, %f8;
	ld.const.f32 	%f10, [kYCbCrOffset+0];
	mul.ftz.f32 	%f11, %f6, %f10;
	sub.ftz.f32 	%f12, %f3, %f11;
	ld.const.f32 	%f13, [kYCbCrOffset+8];
	mul.ftz.f32 	%f14, %f6, %f13;
	sub.ftz.f32 	%f15, %f1, %f14;
	ld.global.f32 	%f16, [k709YCbCr_To_RGB32f+32];
	ld.global.f32 	%f17, [k709YCbCr_To_RGB32f+24];
	ld.global.f32 	%f18, [k709YCbCr_To_RGB32f+28];
	mul.ftz.f32 	%f19, %f9, %f18;
	fma.rn.ftz.f32 	%f20, %f17, %f12, %f19;
	fma.rn.ftz.f32 	%f21, %f16, %f15, %f20;
	ld.global.f32 	%f22, [k709YCbCr_To_RGB32f+20];
	ld.global.f32 	%f23, [k709YCbCr_To_RGB32f+12];
	ld.global.f32 	%f24, [k709YCbCr_To_RGB32f+16];
	mul.ftz.f32 	%f25, %f9, %f24;
	fma.rn.ftz.f32 	%f26, %f23, %f12, %f25;
	fma.rn.ftz.f32 	%f27, %f22, %f15, %f26;
	ld.global.f32 	%f28, [k709YCbCr_To_RGB32f+8];
	ld.global.f32 	%f29, [k709YCbCr_To_RGB32f+0];
	ld.global.f32 	%f30, [k709YCbCr_To_RGB32f+4];
	mul.ftz.f32 	%f31, %f9, %f30;
	fma.rn.ftz.f32 	%f32, %f29, %f12, %f31;
	fma.rn.ftz.f32 	%f33, %f28, %f15, %f32;
	.loc	20	469	0
	cvt.rn.f32.u32 	%f34, %r23;
	add.ftz.f32 	%f35, %f34, %f1;
	cvt.rzi.ftz.u32.f32 	%r24, %f35;
	shr.u32 	%r25, %r24, 1;
	cvt.rn.f32.u32 	%f36, %r25;
	cvt.rn.f32.u32 	%f37, %r22;
	add.ftz.f32 	%f38, %f37, %f2;
	cvt.rzi.ftz.u32.f32 	%r26, %f38;
	shr.u32 	%r27, %r26, 1;
	cvt.rn.f32.u32 	%f39, %r27;
	cvt.rn.f32.u32 	%f40, %r21;
	.loc	20	490	0
	mul.ftz.f32 	%f41, %f6, %f7;
	sub.ftz.f32 	%f42, %f39, %f41;
	mul.ftz.f32 	%f43, %f6, %f10;
	sub.ftz.f32 	%f44, %f40, %f43;
	mul.ftz.f32 	%f45, %f6, %f13;
	sub.ftz.f32 	%f46, %f36, %f45;
	mul.ftz.f32 	%f47, %f42, %f24;
	fma.rn.ftz.f32 	%f48, %f23, %f44, %f47;
	fma.rn.ftz.f32 	%f49, %f22, %f46, %f48;
	mul.ftz.f32 	%f50, %f42, %f18;
	fma.rn.ftz.f32 	%f51, %f17, %f44, %f50;
	fma.rn.ftz.f32 	%f52, %f16, %f46, %f51;
	.loc	21	268	0
	mul.ftz.f32 	%f53, %f42, %f30;
	fma.rn.ftz.f32 	%f54, %f29, %f44, %f53;
	fma.rn.ftz.f32 	%f55, %f28, %f46, %f54;
	.loc	22	200	0
	ld.param.s32 	%r28, [__cudaparm_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDestDeviceFormat];
	mov.s32 	%r29, 0;
	setp.eq.s32 	%p3, %r28, %r29;
	ld.param.s32 	%r30, [__cudaparm_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDestPitch];
	mul.lo.s32 	%r31, %r30, %r5;
	add.s32 	%r32, %r12, %r31;
	cvt.s64.s32 	%rd5, %r32;
	ld.param.u64 	%rd6, [__cudaparm_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDest];
	@!%p3 bra 	$Lt_44_1451010;
	.loc	19	126	0
	mul.lo.u64 	%rd7, %rd5, 8;
	add.u64 	%rd8, %rd6, %rd7;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f21;
	mov.b32		%r33, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f27;
	mov.b32		%r34, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f33;
	mov.b32		%r35, %b1; }
	mov.f32 	%f56, 0f3f800000;    	// 1
	mov.f32 	%f57, 0f437f0000;    	// 255
	div.approx.ftz.f32 	%f58, %f56, %f57;
	mov.f32 	%f59, 0f437f0000;    	// 255
	mul.ftz.f32 	%f60, %f58, %f59;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f60;
	mov.b32		%r36, %b1; }
	st.global.v4.u16 	[%rd8+0], {%r33,%r34,%r35,%r36};
	.loc	22	200	0
	bra.uni 	$Lt_44_1450754;
$Lt_44_1451010:
	.loc	19	126	0
	mul.lo.u64 	%rd9, %rd5, 16;
	add.u64 	%rd10, %rd6, %rd9;
	mov.f32 	%f61, 0f3f800000;    	// 1
	mov.f32 	%f62, 0f437f0000;    	// 255
	div.approx.ftz.f32 	%f63, %f61, %f62;
	mov.f32 	%f64, 0f437f0000;    	// 255
	mul.ftz.f32 	%f60, %f63, %f64;
	st.global.v4.f32 	[%rd10+0], {%f21,%f27,%f33,%f60};
$Lt_44_1450754:
	.loc	22	200	0
	@!%p3 bra 	$Lt_44_1451522;
	.loc	19	126	0
	mul.lo.u64 	%rd11, %rd5, 8;
	add.u64 	%rd8, %rd6, %rd11;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f52;
	mov.b32		%r37, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f49;
	mov.b32		%r38, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f55;
	mov.b32		%r39, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f60;
	mov.b32		%r40, %b1; }
	st.global.v4.u16 	[%rd8+8], {%r37,%r38,%r39,%r40};
	.loc	22	200	0
	bra.uni 	$Lt_44_1476866;
$Lt_44_1451522:
	.loc	19	126	0
	mul.lo.u64 	%rd12, %rd5, 16;
	add.u64 	%rd10, %rd6, %rd12;
	st.global.v4.f32 	[%rd10+16], {%f52,%f49,%f55,%f60};
	bra.uni 	$Lt_44_1476866;
$Lt_44_1435138:
	.loc	22	200	0
	add.s32 	%r41, %r12, 1;
	setp.le.s32 	%p4, %r13, %r41;
	@%p4 bra 	$Lt_44_1452034;
	.loc	19	115	0
	ld.param.u64 	%rd13, [__cudaparm_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inSrc];
	ld.param.s32 	%r42, [__cudaparm_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inSrcPitch];
	mul.lo.s32 	%r43, %r42, %r5;
	add.s32 	%r44, %r11, %r43;
	cvt.s64.s32 	%rd14, %r44;
	mul.wide.s32 	%rd15, %r44, 4;
	add.u64 	%rd16, %rd13, %rd15;
	ld.global.v4.u8 	{%r45,%r46,%r47,%r48}, [%rd16+0];
	.loc	20	469	0
	cvt.rn.f32.u32 	%f65, %r47;
	cvt.rn.f32.u32 	%f66, %r45;
	cvt.rn.f32.u32 	%f67, %r46;
	.loc	22	60	0
	mov.f32 	%f68, 0f437f0000;    	// 255
	mov.f32 	%f69, 0f437f0000;    	// 255
	div.approx.ftz.f32 	%f70, %f68, %f69;
	ld.const.f32 	%f71, [kYCbCrOffset+4];
	mul.ftz.f32 	%f72, %f70, %f71;
	sub.ftz.f32 	%f73, %f66, %f72;
	ld.const.f32 	%f74, [kYCbCrOffset+0];
	mul.ftz.f32 	%f75, %f70, %f74;
	sub.ftz.f32 	%f76, %f67, %f75;
	ld.const.f32 	%f77, [kYCbCrOffset+8];
	mul.ftz.f32 	%f78, %f70, %f77;
	sub.ftz.f32 	%f79, %f65, %f78;
	ld.global.f32 	%f80, [k709YCbCr_To_RGB32f+32];
	ld.global.f32 	%f81, [k709YCbCr_To_RGB32f+24];
	ld.global.f32 	%f82, [k709YCbCr_To_RGB32f+28];
	mul.ftz.f32 	%f83, %f73, %f82;
	fma.rn.ftz.f32 	%f84, %f81, %f76, %f83;
	fma.rn.ftz.f32 	%f85, %f80, %f79, %f84;
	ld.global.f32 	%f86, [k709YCbCr_To_RGB32f+20];
	ld.global.f32 	%f87, [k709YCbCr_To_RGB32f+12];
	ld.global.f32 	%f88, [k709YCbCr_To_RGB32f+16];
	mul.ftz.f32 	%f89, %f73, %f88;
	fma.rn.ftz.f32 	%f90, %f87, %f76, %f89;
	fma.rn.ftz.f32 	%f91, %f86, %f79, %f90;
	ld.global.f32 	%f92, [k709YCbCr_To_RGB32f+8];
	ld.global.f32 	%f93, [k709YCbCr_To_RGB32f+0];
	ld.global.f32 	%f94, [k709YCbCr_To_RGB32f+4];
	mul.ftz.f32 	%f95, %f73, %f94;
	fma.rn.ftz.f32 	%f96, %f93, %f76, %f95;
	fma.rn.ftz.f32 	%f97, %f92, %f79, %f96;
	.loc	20	469	0
	add.ftz.f32 	%f98, %f65, %f65;
	cvt.rzi.ftz.u32.f32 	%r49, %f98;
	shr.u32 	%r50, %r49, 1;
	cvt.rn.f32.u32 	%f99, %r50;
	add.ftz.f32 	%f100, %f66, %f66;
	cvt.rzi.ftz.u32.f32 	%r51, %f100;
	shr.u32 	%r52, %r51, 1;
	cvt.rn.f32.u32 	%f101, %r52;
	cvt.rn.f32.u32 	%f102, %r48;
	.loc	20	490	0
	mul.ftz.f32 	%f103, %f70, %f71;
	sub.ftz.f32 	%f104, %f101, %f103;
	mul.ftz.f32 	%f105, %f70, %f74;
	sub.ftz.f32 	%f106, %f102, %f105;
	mul.ftz.f32 	%f107, %f70, %f77;
	sub.ftz.f32 	%f108, %f99, %f107;
	mul.ftz.f32 	%f109, %f104, %f88;
	fma.rn.ftz.f32 	%f110, %f87, %f106, %f109;
	fma.rn.ftz.f32 	%f111, %f86, %f108, %f110;
	mul.ftz.f32 	%f112, %f104, %f82;
	fma.rn.ftz.f32 	%f113, %f81, %f106, %f112;
	fma.rn.ftz.f32 	%f114, %f80, %f108, %f113;
	.loc	21	268	0
	mul.ftz.f32 	%f115, %f104, %f94;
	fma.rn.ftz.f32 	%f116, %f93, %f106, %f115;
	fma.rn.ftz.f32 	%f117, %f92, %f108, %f116;
	.loc	22	200	0
	ld.param.s32 	%r53, [__cudaparm_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDestDeviceFormat];
	mov.s32 	%r54, 0;
	setp.eq.s32 	%p5, %r53, %r54;
	ld.param.s32 	%r55, [__cudaparm_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDestPitch];
	mul.lo.s32 	%r56, %r55, %r5;
	add.s32 	%r57, %r12, %r56;
	cvt.s64.s32 	%rd17, %r57;
	ld.param.u64 	%rd18, [__cudaparm_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDest];
	@!%p5 bra 	$Lt_44_1467906;
	.loc	19	126	0
	mul.lo.u64 	%rd19, %rd17, 8;
	add.u64 	%rd8, %rd18, %rd19;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f85;
	mov.b32		%r58, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f91;
	mov.b32		%r59, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f97;
	mov.b32		%r60, %b1; }
	mov.f32 	%f118, 0f3f800000;   	// 1
	mov.f32 	%f119, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f120, %f118, %f119;
	mov.f32 	%f121, 0f437f0000;   	// 255
	mul.ftz.f32 	%f60, %f120, %f121;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f60;
	mov.b32		%r61, %b1; }
	st.global.v4.u16 	[%rd8+0], {%r58,%r59,%r60,%r61};
	.loc	22	200	0
	bra.uni 	$Lt_44_1467650;
$Lt_44_1467906:
	.loc	19	126	0
	mul.lo.u64 	%rd20, %rd17, 16;
	add.u64 	%rd10, %rd18, %rd20;
	mov.f32 	%f122, 0f3f800000;   	// 1
	mov.f32 	%f123, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f124, %f122, %f123;
	mov.f32 	%f125, 0f437f0000;   	// 255
	mul.ftz.f32 	%f60, %f124, %f125;
	st.global.v4.f32 	[%rd10+0], {%f85,%f91,%f97,%f60};
$Lt_44_1467650:
	.loc	22	200	0
	@!%p5 bra 	$Lt_44_1468418;
	.loc	19	126	0
	mul.lo.u64 	%rd21, %rd17, 8;
	add.u64 	%rd8, %rd18, %rd21;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f114;
	mov.b32		%r62, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f111;
	mov.b32		%r63, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f117;
	mov.b32		%r64, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f60;
	mov.b32		%r65, %b1; }
	st.global.v4.u16 	[%rd8+8], {%r62,%r63,%r64,%r65};
	.loc	22	200	0
	bra.uni 	$Lt_44_1476866;
$Lt_44_1468418:
	.loc	19	126	0
	mul.lo.u64 	%rd22, %rd17, 16;
	add.u64 	%rd10, %rd18, %rd22;
	st.global.v4.f32 	[%rd10+16], {%f114,%f111,%f117,%f60};
	bra.uni 	$Lt_44_1476866;
$Lt_44_1452034:
	.loc	22	200	0
	setp.ge.s32 	%p6, %r12, %r13;
	@%p6 bra 	$Lt_44_1476866;
	.loc	20	469	0
	ld.param.u64 	%rd23, [__cudaparm_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inSrc];
	ld.param.s32 	%r66, [__cudaparm_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inSrcPitch];
	mul.lo.s32 	%r67, %r66, %r5;
	add.s32 	%r68, %r11, %r67;
	cvt.s64.s32 	%rd24, %r68;
	mul.wide.s32 	%rd25, %r68, 4;
	add.u64 	%rd26, %rd23, %rd25;
	ld.global.v4.u8 	{%r69,%r70,%r71,_}, [%rd26+0];
	cvt.rn.f32.u32 	%f126, %r71;
	cvt.rn.f32.u32 	%f127, %r69;
	cvt.rn.f32.u32 	%f128, %r70;
	.loc	20	490	0
	mov.f32 	%f129, 0f437f0000;   	// 255
	mov.f32 	%f130, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f131, %f129, %f130;
	ld.const.f32 	%f132, [kYCbCrOffset+4];
	mul.ftz.f32 	%f133, %f131, %f132;
	sub.ftz.f32 	%f134, %f127, %f133;
	ld.const.f32 	%f135, [kYCbCrOffset+0];
	mul.ftz.f32 	%f136, %f131, %f135;
	sub.ftz.f32 	%f137, %f128, %f136;
	ld.const.f32 	%f138, [kYCbCrOffset+8];
	mul.ftz.f32 	%f139, %f131, %f138;
	sub.ftz.f32 	%f140, %f126, %f139;
	ld.global.f32 	%f141, [k709YCbCr_To_RGB32f+16];
	mul.ftz.f32 	%f142, %f141, %f134;
	ld.global.f32 	%f143, [k709YCbCr_To_RGB32f+12];
	fma.rn.ftz.f32 	%f144, %f143, %f137, %f142;
	ld.global.f32 	%f145, [k709YCbCr_To_RGB32f+20];
	fma.rn.ftz.f32 	%f146, %f145, %f140, %f144;
	.loc	21	268	0
	ld.global.f32 	%f147, [k709YCbCr_To_RGB32f+4];
	mul.ftz.f32 	%f148, %f147, %f134;
	ld.global.f32 	%f149, [k709YCbCr_To_RGB32f+0];
	fma.rn.ftz.f32 	%f150, %f149, %f137, %f148;
	ld.global.f32 	%f151, [k709YCbCr_To_RGB32f+8];
	fma.rn.ftz.f32 	%f152, %f151, %f140, %f150;
	.loc	22	200	0
	ld.param.s32 	%r72, [__cudaparm_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDestPitch];
	mul.lo.s32 	%r73, %r72, %r5;
	ld.global.f32 	%f153, [k709YCbCr_To_RGB32f+28];
	mul.ftz.f32 	%f154, %f153, %f134;
	add.s32 	%r74, %r12, %r73;
	ld.global.f32 	%f155, [k709YCbCr_To_RGB32f+24];
	fma.rn.ftz.f32 	%f156, %f155, %f137, %f154;
	cvt.s64.s32 	%rd27, %r74;
	ld.global.f32 	%f157, [k709YCbCr_To_RGB32f+32];
	fma.rn.ftz.f32 	%f158, %f157, %f140, %f156;
	ld.param.u64 	%rd28, [__cudaparm_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDest];
	ld.param.s32 	%r75, [__cudaparm_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDestDeviceFormat];
	mov.u32 	%r76, 0;
	setp.ne.s32 	%p7, %r75, %r76;
	@%p7 bra 	$Lt_44_1477122;
	.loc	19	126	0
	mul.lo.u64 	%rd29, %rd27, 8;
	add.u64 	%rd8, %rd28, %rd29;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f158;
	mov.b32		%r77, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f146;
	mov.b32		%r78, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f152;
	mov.b32		%r79, %b1; }
	mov.f32 	%f159, 0f3f800000;   	// 1
	mov.f32 	%f160, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f161, %f159, %f160;
	mov.f32 	%f162, 0f437f0000;   	// 255
	mul.ftz.f32 	%f163, %f161, %f162;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f163;
	mov.b32		%r80, %b1; }
	st.global.v4.u16 	[%rd8+0], {%r77,%r78,%r79,%r80};
	.loc	22	200	0
	bra.uni 	$Lt_44_1476866;
$Lt_44_1477122:
	.loc	19	126	0
	mul.lo.u64 	%rd30, %rd27, 16;
	add.u64 	%rd10, %rd28, %rd30;
	mov.f32 	%f164, 0f3f800000;   	// 1
	mov.f32 	%f165, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f166, %f164, %f165;
	mov.f32 	%f167, 0f437f0000;   	// 255
	mul.ftz.f32 	%f168, %f166, %f167;
	st.global.v4.f32 	[%rd10+0], {%f158,%f146,%f152,%f168};
$Lt_44_1476866:
$Lt_44_1468674:
$Lt_44_1451778:
$Lt_44_1434882:
$Lt_44_1434370:
	.loc	22	200	0
	exit;
$LDWend_PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel:
	} // PixelFormatConvert_IR_PixelFormat_UYVY_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel

	.entry PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel (
		.param .u64 __cudaparm_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inSrc,
		.param .s32 __cudaparm_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inSrcPitch,
		.param .u32 __cudaparm_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel___T2113,
		.param .u64 __cudaparm_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDest,
		.param .s32 __cudaparm_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDestPitch,
		.param .u32 __cudaparm_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDestDeviceFormat,
		.param .s32 __cudaparm_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inWidth,
		.param .s32 __cudaparm_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inHeight)
	{
	.reg .u32 %r<82>;
	.reg .u64 %rd<32>;
	.reg .f32 %f<170>;
	.reg .pred %p<9>;
	.loc	22	201	0
$LDWbegin_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel:
	cvt.s32.u32 	%r1, %ctaid.y;
	cvt.s32.u32 	%r2, %ntid.y;
	mul.lo.s32 	%r3, %r1, %r2;
	mov.u32 	%r4, %tid.y;
	add.u32 	%r5, %r3, %r4;
	ld.param.s32 	%r6, [__cudaparm_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inHeight];
	setp.le.s32 	%p1, %r6, %r5;
	@%p1 bra 	$Lt_45_1475586;
	cvt.s32.u32 	%r7, %ctaid.x;
	cvt.s32.u32 	%r8, %ntid.x;
	mul.lo.s32 	%r9, %r7, %r8;
	mov.u32 	%r10, %tid.x;
	add.u32 	%r11, %r9, %r10;
	shl.b32 	%r12, %r11, 1;
	ld.param.s32 	%r13, [__cudaparm_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inWidth];
	add.s32 	%r14, %r12, 2;
	setp.le.s32 	%p2, %r13, %r14;
	@%p2 bra 	$Lt_45_1433858;
	.loc	19	115	0
	ld.param.u64 	%rd1, [__cudaparm_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inSrc];
	ld.param.s32 	%r15, [__cudaparm_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inSrcPitch];
	mul.lo.s32 	%r16, %r15, %r5;
	add.s32 	%r17, %r11, %r16;
	cvt.s64.s32 	%rd2, %r17;
	mul.wide.s32 	%rd3, %r17, 4;
	add.u64 	%rd4, %rd1, %rd3;
	ld.global.v4.u8 	{%r18,%r19,%r20,%r21}, [%rd4+0];
	ld.global.u8 	%r22, [%rd4+5];
	ld.global.u8 	%r23, [%rd4+7];
	.loc	20	469	0
	cvt.rn.f32.u32 	%f1, %r21;
	cvt.rn.f32.u32 	%f2, %r19;
	cvt.rn.f32.u32 	%f3, %r18;
	.loc	22	60	0
	mov.f32 	%f4, 0f437f0000;     	// 255
	mov.f32 	%f5, 0f437f0000;     	// 255
	div.approx.ftz.f32 	%f6, %f4, %f5;
	ld.const.f32 	%f7, [kYCbCrOffset+4];
	mul.ftz.f32 	%f8, %f6, %f7;
	sub.ftz.f32 	%f9, %f2, %f8;
	ld.const.f32 	%f10, [kYCbCrOffset+0];
	mul.ftz.f32 	%f11, %f6, %f10;
	sub.ftz.f32 	%f12, %f3, %f11;
	ld.const.f32 	%f13, [kYCbCrOffset+8];
	mul.ftz.f32 	%f14, %f6, %f13;
	sub.ftz.f32 	%f15, %f1, %f14;
	ld.global.f32 	%f16, [k601YCbCr_To_RGB32f+32];
	ld.global.f32 	%f17, [k601YCbCr_To_RGB32f+24];
	ld.global.f32 	%f18, [k601YCbCr_To_RGB32f+28];
	mul.ftz.f32 	%f19, %f9, %f18;
	fma.rn.ftz.f32 	%f20, %f17, %f12, %f19;
	fma.rn.ftz.f32 	%f21, %f16, %f15, %f20;
	ld.global.f32 	%f22, [k601YCbCr_To_RGB32f+20];
	ld.global.f32 	%f23, [k601YCbCr_To_RGB32f+12];
	ld.global.f32 	%f24, [k601YCbCr_To_RGB32f+16];
	mul.ftz.f32 	%f25, %f9, %f24;
	fma.rn.ftz.f32 	%f26, %f23, %f12, %f25;
	fma.rn.ftz.f32 	%f27, %f22, %f15, %f26;
	ld.global.f32 	%f28, [k601YCbCr_To_RGB32f+8];
	ld.global.f32 	%f29, [k601YCbCr_To_RGB32f+0];
	ld.global.f32 	%f30, [k601YCbCr_To_RGB32f+4];
	mul.ftz.f32 	%f31, %f9, %f30;
	fma.rn.ftz.f32 	%f32, %f29, %f12, %f31;
	fma.rn.ftz.f32 	%f33, %f28, %f15, %f32;
	.loc	20	469	0
	cvt.rn.f32.u32 	%f34, %r23;
	add.ftz.f32 	%f35, %f34, %f1;
	cvt.rzi.ftz.u32.f32 	%r24, %f35;
	shr.u32 	%r25, %r24, 1;
	cvt.rn.f32.u32 	%f36, %r25;
	cvt.rn.f32.u32 	%f37, %r22;
	add.ftz.f32 	%f38, %f37, %f2;
	cvt.rzi.ftz.u32.f32 	%r26, %f38;
	shr.u32 	%r27, %r26, 1;
	cvt.rn.f32.u32 	%f39, %r27;
	cvt.rn.f32.u32 	%f40, %r20;
	.loc	20	490	0
	mul.ftz.f32 	%f41, %f6, %f7;
	sub.ftz.f32 	%f42, %f39, %f41;
	mul.ftz.f32 	%f43, %f6, %f10;
	sub.ftz.f32 	%f44, %f40, %f43;
	mul.ftz.f32 	%f45, %f6, %f13;
	sub.ftz.f32 	%f46, %f36, %f45;
	mul.ftz.f32 	%f47, %f42, %f24;
	fma.rn.ftz.f32 	%f48, %f23, %f44, %f47;
	fma.rn.ftz.f32 	%f49, %f22, %f46, %f48;
	mul.ftz.f32 	%f50, %f42, %f18;
	fma.rn.ftz.f32 	%f51, %f17, %f44, %f50;
	fma.rn.ftz.f32 	%f52, %f16, %f46, %f51;
	.loc	21	268	0
	mul.ftz.f32 	%f53, %f42, %f30;
	fma.rn.ftz.f32 	%f54, %f29, %f44, %f53;
	fma.rn.ftz.f32 	%f55, %f28, %f46, %f54;
	.loc	22	201	0
	ld.param.s32 	%r28, [__cudaparm_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDestDeviceFormat];
	mov.s32 	%r29, 0;
	setp.eq.s32 	%p3, %r28, %r29;
	ld.param.s32 	%r30, [__cudaparm_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDestPitch];
	mul.lo.s32 	%r31, %r30, %r5;
	add.s32 	%r32, %r12, %r31;
	cvt.s64.s32 	%rd5, %r32;
	ld.param.u64 	%rd6, [__cudaparm_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDest];
	@!%p3 bra 	$Lt_45_1449730;
	.loc	19	126	0
	mul.lo.u64 	%rd7, %rd5, 8;
	add.u64 	%rd8, %rd6, %rd7;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f21;
	mov.b32		%r33, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f27;
	mov.b32		%r34, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f33;
	mov.b32		%r35, %b1; }
	mov.f32 	%f56, 0f3f800000;    	// 1
	mov.f32 	%f57, 0f437f0000;    	// 255
	div.approx.ftz.f32 	%f58, %f56, %f57;
	mov.f32 	%f59, 0f437f0000;    	// 255
	mul.ftz.f32 	%f60, %f58, %f59;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f60;
	mov.b32		%r36, %b1; }
	st.global.v4.u16 	[%rd8+0], {%r33,%r34,%r35,%r36};
	.loc	22	201	0
	bra.uni 	$Lt_45_1449474;
$Lt_45_1449730:
	.loc	19	126	0
	mul.lo.u64 	%rd9, %rd5, 16;
	add.u64 	%rd10, %rd6, %rd9;
	mov.f32 	%f61, 0f3f800000;    	// 1
	mov.f32 	%f62, 0f437f0000;    	// 255
	div.approx.ftz.f32 	%f63, %f61, %f62;
	mov.f32 	%f64, 0f437f0000;    	// 255
	mul.ftz.f32 	%f60, %f63, %f64;
	st.global.v4.f32 	[%rd10+0], {%f21,%f27,%f33,%f60};
$Lt_45_1449474:
	.loc	22	201	0
	@!%p3 bra 	$Lt_45_1450242;
	.loc	19	126	0
	mul.lo.u64 	%rd11, %rd5, 8;
	add.u64 	%rd8, %rd6, %rd11;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f52;
	mov.b32		%r37, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f49;
	mov.b32		%r38, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f55;
	mov.b32		%r39, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f60;
	mov.b32		%r40, %b1; }
	st.global.v4.u16 	[%rd8+8], {%r37,%r38,%r39,%r40};
	.loc	22	201	0
	bra.uni 	$Lt_45_1475586;
$Lt_45_1450242:
	.loc	19	126	0
	mul.lo.u64 	%rd12, %rd5, 16;
	add.u64 	%rd10, %rd6, %rd12;
	st.global.v4.f32 	[%rd10+16], {%f52,%f49,%f55,%f60};
	bra.uni 	$Lt_45_1475586;
$Lt_45_1433858:
	.loc	22	201	0
	add.s32 	%r41, %r12, 1;
	setp.le.s32 	%p4, %r13, %r41;
	@%p4 bra 	$Lt_45_1450754;
	.loc	19	115	0
	ld.param.u64 	%rd13, [__cudaparm_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inSrc];
	ld.param.s32 	%r42, [__cudaparm_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inSrcPitch];
	mul.lo.s32 	%r43, %r42, %r5;
	add.s32 	%r44, %r11, %r43;
	cvt.s64.s32 	%rd14, %r44;
	mul.wide.s32 	%rd15, %r44, 4;
	add.u64 	%rd16, %rd13, %rd15;
	ld.global.v4.u8 	{%r45,%r46,%r47,%r48}, [%rd16+0];
	.loc	20	469	0
	cvt.rn.f32.u32 	%f65, %r48;
	cvt.rn.f32.u32 	%f66, %r46;
	cvt.rn.f32.u32 	%f67, %r45;
	.loc	22	60	0
	mov.f32 	%f68, 0f437f0000;    	// 255
	mov.f32 	%f69, 0f437f0000;    	// 255
	div.approx.ftz.f32 	%f70, %f68, %f69;
	ld.const.f32 	%f71, [kYCbCrOffset+4];
	mul.ftz.f32 	%f72, %f70, %f71;
	sub.ftz.f32 	%f73, %f66, %f72;
	ld.const.f32 	%f74, [kYCbCrOffset+0];
	mul.ftz.f32 	%f75, %f70, %f74;
	sub.ftz.f32 	%f76, %f67, %f75;
	ld.const.f32 	%f77, [kYCbCrOffset+8];
	mul.ftz.f32 	%f78, %f70, %f77;
	sub.ftz.f32 	%f79, %f65, %f78;
	ld.global.f32 	%f80, [k601YCbCr_To_RGB32f+32];
	ld.global.f32 	%f81, [k601YCbCr_To_RGB32f+24];
	ld.global.f32 	%f82, [k601YCbCr_To_RGB32f+28];
	mul.ftz.f32 	%f83, %f73, %f82;
	fma.rn.ftz.f32 	%f84, %f81, %f76, %f83;
	fma.rn.ftz.f32 	%f85, %f80, %f79, %f84;
	ld.global.f32 	%f86, [k601YCbCr_To_RGB32f+20];
	ld.global.f32 	%f87, [k601YCbCr_To_RGB32f+12];
	ld.global.f32 	%f88, [k601YCbCr_To_RGB32f+16];
	mul.ftz.f32 	%f89, %f73, %f88;
	fma.rn.ftz.f32 	%f90, %f87, %f76, %f89;
	fma.rn.ftz.f32 	%f91, %f86, %f79, %f90;
	ld.global.f32 	%f92, [k601YCbCr_To_RGB32f+8];
	ld.global.f32 	%f93, [k601YCbCr_To_RGB32f+0];
	ld.global.f32 	%f94, [k601YCbCr_To_RGB32f+4];
	mul.ftz.f32 	%f95, %f73, %f94;
	fma.rn.ftz.f32 	%f96, %f93, %f76, %f95;
	fma.rn.ftz.f32 	%f97, %f92, %f79, %f96;
	.loc	20	469	0
	add.ftz.f32 	%f98, %f65, %f65;
	cvt.rzi.ftz.u32.f32 	%r49, %f98;
	shr.u32 	%r50, %r49, 1;
	cvt.rn.f32.u32 	%f99, %r50;
	add.ftz.f32 	%f100, %f66, %f66;
	cvt.rzi.ftz.u32.f32 	%r51, %f100;
	shr.u32 	%r52, %r51, 1;
	cvt.rn.f32.u32 	%f101, %r52;
	cvt.rn.f32.u32 	%f102, %r47;
	.loc	20	490	0
	mul.ftz.f32 	%f103, %f70, %f71;
	sub.ftz.f32 	%f104, %f101, %f103;
	mul.ftz.f32 	%f105, %f70, %f74;
	sub.ftz.f32 	%f106, %f102, %f105;
	mul.ftz.f32 	%f107, %f70, %f77;
	sub.ftz.f32 	%f108, %f99, %f107;
	mul.ftz.f32 	%f109, %f104, %f88;
	fma.rn.ftz.f32 	%f110, %f87, %f106, %f109;
	fma.rn.ftz.f32 	%f111, %f86, %f108, %f110;
	mul.ftz.f32 	%f112, %f104, %f82;
	fma.rn.ftz.f32 	%f113, %f81, %f106, %f112;
	fma.rn.ftz.f32 	%f114, %f80, %f108, %f113;
	.loc	21	268	0
	mul.ftz.f32 	%f115, %f104, %f94;
	fma.rn.ftz.f32 	%f116, %f93, %f106, %f115;
	fma.rn.ftz.f32 	%f117, %f92, %f108, %f116;
	.loc	22	201	0
	ld.param.s32 	%r53, [__cudaparm_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDestDeviceFormat];
	mov.s32 	%r54, 0;
	setp.eq.s32 	%p5, %r53, %r54;
	ld.param.s32 	%r55, [__cudaparm_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDestPitch];
	mul.lo.s32 	%r56, %r55, %r5;
	add.s32 	%r57, %r12, %r56;
	cvt.s64.s32 	%rd17, %r57;
	ld.param.u64 	%rd18, [__cudaparm_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDest];
	@!%p5 bra 	$Lt_45_1466626;
	.loc	19	126	0
	mul.lo.u64 	%rd19, %rd17, 8;
	add.u64 	%rd8, %rd18, %rd19;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f85;
	mov.b32		%r58, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f91;
	mov.b32		%r59, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f97;
	mov.b32		%r60, %b1; }
	mov.f32 	%f118, 0f3f800000;   	// 1
	mov.f32 	%f119, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f120, %f118, %f119;
	mov.f32 	%f121, 0f437f0000;   	// 255
	mul.ftz.f32 	%f60, %f120, %f121;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f60;
	mov.b32		%r61, %b1; }
	st.global.v4.u16 	[%rd8+0], {%r58,%r59,%r60,%r61};
	.loc	22	201	0
	bra.uni 	$Lt_45_1466370;
$Lt_45_1466626:
	.loc	19	126	0
	mul.lo.u64 	%rd20, %rd17, 16;
	add.u64 	%rd10, %rd18, %rd20;
	mov.f32 	%f122, 0f3f800000;   	// 1
	mov.f32 	%f123, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f124, %f122, %f123;
	mov.f32 	%f125, 0f437f0000;   	// 255
	mul.ftz.f32 	%f60, %f124, %f125;
	st.global.v4.f32 	[%rd10+0], {%f85,%f91,%f97,%f60};
$Lt_45_1466370:
	.loc	22	201	0
	@!%p5 bra 	$Lt_45_1467138;
	.loc	19	126	0
	mul.lo.u64 	%rd21, %rd17, 8;
	add.u64 	%rd8, %rd18, %rd21;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f114;
	mov.b32		%r62, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f111;
	mov.b32		%r63, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f117;
	mov.b32		%r64, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f60;
	mov.b32		%r65, %b1; }
	st.global.v4.u16 	[%rd8+8], {%r62,%r63,%r64,%r65};
	.loc	22	201	0
	bra.uni 	$Lt_45_1475586;
$Lt_45_1467138:
	.loc	19	126	0
	mul.lo.u64 	%rd22, %rd17, 16;
	add.u64 	%rd10, %rd18, %rd22;
	st.global.v4.f32 	[%rd10+16], {%f114,%f111,%f117,%f60};
	bra.uni 	$Lt_45_1475586;
$Lt_45_1450754:
	.loc	22	201	0
	setp.ge.s32 	%p6, %r12, %r13;
	@%p6 bra 	$Lt_45_1475586;
	.loc	20	469	0
	ld.param.u64 	%rd23, [__cudaparm_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inSrc];
	ld.param.s32 	%r66, [__cudaparm_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inSrcPitch];
	mul.lo.s32 	%r67, %r66, %r5;
	add.s32 	%r68, %r11, %r67;
	cvt.s64.s32 	%rd24, %r68;
	mul.wide.s32 	%rd25, %r68, 4;
	add.u64 	%rd26, %rd23, %rd25;
	ld.global.v4.u8 	{%r69,%r70,_,%r71}, [%rd26+0];
	cvt.rn.f32.u32 	%f126, %r71;
	cvt.rn.f32.u32 	%f127, %r70;
	cvt.rn.f32.u32 	%f128, %r69;
	.loc	20	490	0
	mov.f32 	%f129, 0f437f0000;   	// 255
	mov.f32 	%f130, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f131, %f129, %f130;
	ld.const.f32 	%f132, [kYCbCrOffset+4];
	mul.ftz.f32 	%f133, %f131, %f132;
	sub.ftz.f32 	%f134, %f127, %f133;
	ld.const.f32 	%f135, [kYCbCrOffset+0];
	mul.ftz.f32 	%f136, %f131, %f135;
	sub.ftz.f32 	%f137, %f128, %f136;
	ld.const.f32 	%f138, [kYCbCrOffset+8];
	mul.ftz.f32 	%f139, %f131, %f138;
	sub.ftz.f32 	%f140, %f126, %f139;
	ld.global.f32 	%f141, [k601YCbCr_To_RGB32f+16];
	mul.ftz.f32 	%f142, %f141, %f134;
	ld.global.f32 	%f143, [k601YCbCr_To_RGB32f+12];
	fma.rn.ftz.f32 	%f144, %f143, %f137, %f142;
	ld.global.f32 	%f145, [k601YCbCr_To_RGB32f+20];
	fma.rn.ftz.f32 	%f146, %f145, %f140, %f144;
	.loc	21	268	0
	ld.global.f32 	%f147, [k601YCbCr_To_RGB32f+4];
	mul.ftz.f32 	%f148, %f147, %f134;
	ld.global.f32 	%f149, [k601YCbCr_To_RGB32f+0];
	fma.rn.ftz.f32 	%f150, %f149, %f137, %f148;
	ld.global.f32 	%f151, [k601YCbCr_To_RGB32f+8];
	fma.rn.ftz.f32 	%f152, %f151, %f140, %f150;
	.loc	22	201	0
	ld.param.s32 	%r72, [__cudaparm_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDestPitch];
	mul.lo.s32 	%r73, %r72, %r5;
	ld.global.f32 	%f153, [k601YCbCr_To_RGB32f+28];
	mul.ftz.f32 	%f154, %f153, %f134;
	add.s32 	%r74, %r12, %r73;
	ld.global.f32 	%f155, [k601YCbCr_To_RGB32f+24];
	fma.rn.ftz.f32 	%f156, %f155, %f137, %f154;
	cvt.s64.s32 	%rd27, %r74;
	ld.global.f32 	%f157, [k601YCbCr_To_RGB32f+32];
	fma.rn.ftz.f32 	%f158, %f157, %f140, %f156;
	ld.param.u64 	%rd28, [__cudaparm_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDest];
	ld.param.s32 	%r75, [__cudaparm_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDestDeviceFormat];
	mov.u32 	%r76, 0;
	setp.ne.s32 	%p7, %r75, %r76;
	@%p7 bra 	$Lt_45_1475842;
	.loc	19	126	0
	mul.lo.u64 	%rd29, %rd27, 8;
	add.u64 	%rd8, %rd28, %rd29;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f158;
	mov.b32		%r77, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f146;
	mov.b32		%r78, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f152;
	mov.b32		%r79, %b1; }
	mov.f32 	%f159, 0f3f800000;   	// 1
	mov.f32 	%f160, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f161, %f159, %f160;
	mov.f32 	%f162, 0f437f0000;   	// 255
	mul.ftz.f32 	%f163, %f161, %f162;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f163;
	mov.b32		%r80, %b1; }
	st.global.v4.u16 	[%rd8+0], {%r77,%r78,%r79,%r80};
	.loc	22	201	0
	bra.uni 	$Lt_45_1475586;
$Lt_45_1475842:
	.loc	19	126	0
	mul.lo.u64 	%rd30, %rd27, 16;
	add.u64 	%rd10, %rd28, %rd30;
	mov.f32 	%f164, 0f3f800000;   	// 1
	mov.f32 	%f165, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f166, %f164, %f165;
	mov.f32 	%f167, 0f437f0000;   	// 255
	mul.ftz.f32 	%f168, %f166, %f167;
	st.global.v4.f32 	[%rd10+0], {%f158,%f146,%f152,%f168};
$Lt_45_1475586:
$Lt_45_1467394:
$Lt_45_1450498:
$Lt_45_1433602:
$Lt_45_1433090:
	.loc	22	201	0
	exit;
$LDWend_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel:
	} // PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_To_IR_PixelFormat_BGRA_4444_32f_Kernel

	.entry PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel (
		.param .u64 __cudaparm_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inSrc,
		.param .s32 __cudaparm_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inSrcPitch,
		.param .u32 __cudaparm_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel___T2125,
		.param .u64 __cudaparm_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDest,
		.param .s32 __cudaparm_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDestPitch,
		.param .u32 __cudaparm_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDestDeviceFormat,
		.param .s32 __cudaparm_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inWidth,
		.param .s32 __cudaparm_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inHeight)
	{
	.reg .u32 %r<82>;
	.reg .u64 %rd<32>;
	.reg .f32 %f<170>;
	.reg .pred %p<9>;
	.loc	22	202	0
$LDWbegin_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel:
	cvt.s32.u32 	%r1, %ctaid.y;
	cvt.s32.u32 	%r2, %ntid.y;
	mul.lo.s32 	%r3, %r1, %r2;
	mov.u32 	%r4, %tid.y;
	add.u32 	%r5, %r3, %r4;
	ld.param.s32 	%r6, [__cudaparm_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inHeight];
	setp.le.s32 	%p1, %r6, %r5;
	@%p1 bra 	$Lt_46_1475586;
	cvt.s32.u32 	%r7, %ctaid.x;
	cvt.s32.u32 	%r8, %ntid.x;
	mul.lo.s32 	%r9, %r7, %r8;
	mov.u32 	%r10, %tid.x;
	add.u32 	%r11, %r9, %r10;
	shl.b32 	%r12, %r11, 1;
	ld.param.s32 	%r13, [__cudaparm_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inWidth];
	add.s32 	%r14, %r12, 2;
	setp.le.s32 	%p2, %r13, %r14;
	@%p2 bra 	$Lt_46_1433858;
	.loc	19	115	0
	ld.param.u64 	%rd1, [__cudaparm_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inSrc];
	ld.param.s32 	%r15, [__cudaparm_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inSrcPitch];
	mul.lo.s32 	%r16, %r15, %r5;
	add.s32 	%r17, %r11, %r16;
	cvt.s64.s32 	%rd2, %r17;
	mul.wide.s32 	%rd3, %r17, 4;
	add.u64 	%rd4, %rd1, %rd3;
	ld.global.v4.u8 	{%r18,%r19,%r20,%r21}, [%rd4+0];
	ld.global.u8 	%r22, [%rd4+5];
	ld.global.u8 	%r23, [%rd4+7];
	.loc	20	469	0
	cvt.rn.f32.u32 	%f1, %r21;
	cvt.rn.f32.u32 	%f2, %r19;
	cvt.rn.f32.u32 	%f3, %r18;
	.loc	22	60	0
	mov.f32 	%f4, 0f437f0000;     	// 255
	mov.f32 	%f5, 0f437f0000;     	// 255
	div.approx.ftz.f32 	%f6, %f4, %f5;
	ld.const.f32 	%f7, [kYCbCrOffset+4];
	mul.ftz.f32 	%f8, %f6, %f7;
	sub.ftz.f32 	%f9, %f2, %f8;
	ld.const.f32 	%f10, [kYCbCrOffset+0];
	mul.ftz.f32 	%f11, %f6, %f10;
	sub.ftz.f32 	%f12, %f3, %f11;
	ld.const.f32 	%f13, [kYCbCrOffset+8];
	mul.ftz.f32 	%f14, %f6, %f13;
	sub.ftz.f32 	%f15, %f1, %f14;
	ld.global.f32 	%f16, [k709YCbCr_To_RGB32f+32];
	ld.global.f32 	%f17, [k709YCbCr_To_RGB32f+24];
	ld.global.f32 	%f18, [k709YCbCr_To_RGB32f+28];
	mul.ftz.f32 	%f19, %f9, %f18;
	fma.rn.ftz.f32 	%f20, %f17, %f12, %f19;
	fma.rn.ftz.f32 	%f21, %f16, %f15, %f20;
	ld.global.f32 	%f22, [k709YCbCr_To_RGB32f+20];
	ld.global.f32 	%f23, [k709YCbCr_To_RGB32f+12];
	ld.global.f32 	%f24, [k709YCbCr_To_RGB32f+16];
	mul.ftz.f32 	%f25, %f9, %f24;
	fma.rn.ftz.f32 	%f26, %f23, %f12, %f25;
	fma.rn.ftz.f32 	%f27, %f22, %f15, %f26;
	ld.global.f32 	%f28, [k709YCbCr_To_RGB32f+8];
	ld.global.f32 	%f29, [k709YCbCr_To_RGB32f+0];
	ld.global.f32 	%f30, [k709YCbCr_To_RGB32f+4];
	mul.ftz.f32 	%f31, %f9, %f30;
	fma.rn.ftz.f32 	%f32, %f29, %f12, %f31;
	fma.rn.ftz.f32 	%f33, %f28, %f15, %f32;
	.loc	20	469	0
	cvt.rn.f32.u32 	%f34, %r23;
	add.ftz.f32 	%f35, %f34, %f1;
	cvt.rzi.ftz.u32.f32 	%r24, %f35;
	shr.u32 	%r25, %r24, 1;
	cvt.rn.f32.u32 	%f36, %r25;
	cvt.rn.f32.u32 	%f37, %r22;
	add.ftz.f32 	%f38, %f37, %f2;
	cvt.rzi.ftz.u32.f32 	%r26, %f38;
	shr.u32 	%r27, %r26, 1;
	cvt.rn.f32.u32 	%f39, %r27;
	cvt.rn.f32.u32 	%f40, %r20;
	.loc	20	490	0
	mul.ftz.f32 	%f41, %f6, %f7;
	sub.ftz.f32 	%f42, %f39, %f41;
	mul.ftz.f32 	%f43, %f6, %f10;
	sub.ftz.f32 	%f44, %f40, %f43;
	mul.ftz.f32 	%f45, %f6, %f13;
	sub.ftz.f32 	%f46, %f36, %f45;
	mul.ftz.f32 	%f47, %f42, %f24;
	fma.rn.ftz.f32 	%f48, %f23, %f44, %f47;
	fma.rn.ftz.f32 	%f49, %f22, %f46, %f48;
	mul.ftz.f32 	%f50, %f42, %f18;
	fma.rn.ftz.f32 	%f51, %f17, %f44, %f50;
	fma.rn.ftz.f32 	%f52, %f16, %f46, %f51;
	.loc	21	268	0
	mul.ftz.f32 	%f53, %f42, %f30;
	fma.rn.ftz.f32 	%f54, %f29, %f44, %f53;
	fma.rn.ftz.f32 	%f55, %f28, %f46, %f54;
	.loc	22	202	0
	ld.param.s32 	%r28, [__cudaparm_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDestDeviceFormat];
	mov.s32 	%r29, 0;
	setp.eq.s32 	%p3, %r28, %r29;
	ld.param.s32 	%r30, [__cudaparm_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDestPitch];
	mul.lo.s32 	%r31, %r30, %r5;
	add.s32 	%r32, %r12, %r31;
	cvt.s64.s32 	%rd5, %r32;
	ld.param.u64 	%rd6, [__cudaparm_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDest];
	@!%p3 bra 	$Lt_46_1449730;
	.loc	19	126	0
	mul.lo.u64 	%rd7, %rd5, 8;
	add.u64 	%rd8, %rd6, %rd7;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f21;
	mov.b32		%r33, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f27;
	mov.b32		%r34, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f33;
	mov.b32		%r35, %b1; }
	mov.f32 	%f56, 0f3f800000;    	// 1
	mov.f32 	%f57, 0f437f0000;    	// 255
	div.approx.ftz.f32 	%f58, %f56, %f57;
	mov.f32 	%f59, 0f437f0000;    	// 255
	mul.ftz.f32 	%f60, %f58, %f59;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f60;
	mov.b32		%r36, %b1; }
	st.global.v4.u16 	[%rd8+0], {%r33,%r34,%r35,%r36};
	.loc	22	202	0
	bra.uni 	$Lt_46_1449474;
$Lt_46_1449730:
	.loc	19	126	0
	mul.lo.u64 	%rd9, %rd5, 16;
	add.u64 	%rd10, %rd6, %rd9;
	mov.f32 	%f61, 0f3f800000;    	// 1
	mov.f32 	%f62, 0f437f0000;    	// 255
	div.approx.ftz.f32 	%f63, %f61, %f62;
	mov.f32 	%f64, 0f437f0000;    	// 255
	mul.ftz.f32 	%f60, %f63, %f64;
	st.global.v4.f32 	[%rd10+0], {%f21,%f27,%f33,%f60};
$Lt_46_1449474:
	.loc	22	202	0
	@!%p3 bra 	$Lt_46_1450242;
	.loc	19	126	0
	mul.lo.u64 	%rd11, %rd5, 8;
	add.u64 	%rd8, %rd6, %rd11;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f52;
	mov.b32		%r37, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f49;
	mov.b32		%r38, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f55;
	mov.b32		%r39, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f60;
	mov.b32		%r40, %b1; }
	st.global.v4.u16 	[%rd8+8], {%r37,%r38,%r39,%r40};
	.loc	22	202	0
	bra.uni 	$Lt_46_1475586;
$Lt_46_1450242:
	.loc	19	126	0
	mul.lo.u64 	%rd12, %rd5, 16;
	add.u64 	%rd10, %rd6, %rd12;
	st.global.v4.f32 	[%rd10+16], {%f52,%f49,%f55,%f60};
	bra.uni 	$Lt_46_1475586;
$Lt_46_1433858:
	.loc	22	202	0
	add.s32 	%r41, %r12, 1;
	setp.le.s32 	%p4, %r13, %r41;
	@%p4 bra 	$Lt_46_1450754;
	.loc	19	115	0
	ld.param.u64 	%rd13, [__cudaparm_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inSrc];
	ld.param.s32 	%r42, [__cudaparm_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inSrcPitch];
	mul.lo.s32 	%r43, %r42, %r5;
	add.s32 	%r44, %r11, %r43;
	cvt.s64.s32 	%rd14, %r44;
	mul.wide.s32 	%rd15, %r44, 4;
	add.u64 	%rd16, %rd13, %rd15;
	ld.global.v4.u8 	{%r45,%r46,%r47,%r48}, [%rd16+0];
	.loc	20	469	0
	cvt.rn.f32.u32 	%f65, %r48;
	cvt.rn.f32.u32 	%f66, %r46;
	cvt.rn.f32.u32 	%f67, %r45;
	.loc	22	60	0
	mov.f32 	%f68, 0f437f0000;    	// 255
	mov.f32 	%f69, 0f437f0000;    	// 255
	div.approx.ftz.f32 	%f70, %f68, %f69;
	ld.const.f32 	%f71, [kYCbCrOffset+4];
	mul.ftz.f32 	%f72, %f70, %f71;
	sub.ftz.f32 	%f73, %f66, %f72;
	ld.const.f32 	%f74, [kYCbCrOffset+0];
	mul.ftz.f32 	%f75, %f70, %f74;
	sub.ftz.f32 	%f76, %f67, %f75;
	ld.const.f32 	%f77, [kYCbCrOffset+8];
	mul.ftz.f32 	%f78, %f70, %f77;
	sub.ftz.f32 	%f79, %f65, %f78;
	ld.global.f32 	%f80, [k709YCbCr_To_RGB32f+32];
	ld.global.f32 	%f81, [k709YCbCr_To_RGB32f+24];
	ld.global.f32 	%f82, [k709YCbCr_To_RGB32f+28];
	mul.ftz.f32 	%f83, %f73, %f82;
	fma.rn.ftz.f32 	%f84, %f81, %f76, %f83;
	fma.rn.ftz.f32 	%f85, %f80, %f79, %f84;
	ld.global.f32 	%f86, [k709YCbCr_To_RGB32f+20];
	ld.global.f32 	%f87, [k709YCbCr_To_RGB32f+12];
	ld.global.f32 	%f88, [k709YCbCr_To_RGB32f+16];
	mul.ftz.f32 	%f89, %f73, %f88;
	fma.rn.ftz.f32 	%f90, %f87, %f76, %f89;
	fma.rn.ftz.f32 	%f91, %f86, %f79, %f90;
	ld.global.f32 	%f92, [k709YCbCr_To_RGB32f+8];
	ld.global.f32 	%f93, [k709YCbCr_To_RGB32f+0];
	ld.global.f32 	%f94, [k709YCbCr_To_RGB32f+4];
	mul.ftz.f32 	%f95, %f73, %f94;
	fma.rn.ftz.f32 	%f96, %f93, %f76, %f95;
	fma.rn.ftz.f32 	%f97, %f92, %f79, %f96;
	.loc	20	469	0
	add.ftz.f32 	%f98, %f65, %f65;
	cvt.rzi.ftz.u32.f32 	%r49, %f98;
	shr.u32 	%r50, %r49, 1;
	cvt.rn.f32.u32 	%f99, %r50;
	add.ftz.f32 	%f100, %f66, %f66;
	cvt.rzi.ftz.u32.f32 	%r51, %f100;
	shr.u32 	%r52, %r51, 1;
	cvt.rn.f32.u32 	%f101, %r52;
	cvt.rn.f32.u32 	%f102, %r47;
	.loc	20	490	0
	mul.ftz.f32 	%f103, %f70, %f71;
	sub.ftz.f32 	%f104, %f101, %f103;
	mul.ftz.f32 	%f105, %f70, %f74;
	sub.ftz.f32 	%f106, %f102, %f105;
	mul.ftz.f32 	%f107, %f70, %f77;
	sub.ftz.f32 	%f108, %f99, %f107;
	mul.ftz.f32 	%f109, %f104, %f88;
	fma.rn.ftz.f32 	%f110, %f87, %f106, %f109;
	fma.rn.ftz.f32 	%f111, %f86, %f108, %f110;
	mul.ftz.f32 	%f112, %f104, %f82;
	fma.rn.ftz.f32 	%f113, %f81, %f106, %f112;
	fma.rn.ftz.f32 	%f114, %f80, %f108, %f113;
	.loc	21	268	0
	mul.ftz.f32 	%f115, %f104, %f94;
	fma.rn.ftz.f32 	%f116, %f93, %f106, %f115;
	fma.rn.ftz.f32 	%f117, %f92, %f108, %f116;
	.loc	22	202	0
	ld.param.s32 	%r53, [__cudaparm_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDestDeviceFormat];
	mov.s32 	%r54, 0;
	setp.eq.s32 	%p5, %r53, %r54;
	ld.param.s32 	%r55, [__cudaparm_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDestPitch];
	mul.lo.s32 	%r56, %r55, %r5;
	add.s32 	%r57, %r12, %r56;
	cvt.s64.s32 	%rd17, %r57;
	ld.param.u64 	%rd18, [__cudaparm_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDest];
	@!%p5 bra 	$Lt_46_1466626;
	.loc	19	126	0
	mul.lo.u64 	%rd19, %rd17, 8;
	add.u64 	%rd8, %rd18, %rd19;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f85;
	mov.b32		%r58, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f91;
	mov.b32		%r59, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f97;
	mov.b32		%r60, %b1; }
	mov.f32 	%f118, 0f3f800000;   	// 1
	mov.f32 	%f119, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f120, %f118, %f119;
	mov.f32 	%f121, 0f437f0000;   	// 255
	mul.ftz.f32 	%f60, %f120, %f121;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f60;
	mov.b32		%r61, %b1; }
	st.global.v4.u16 	[%rd8+0], {%r58,%r59,%r60,%r61};
	.loc	22	202	0
	bra.uni 	$Lt_46_1466370;
$Lt_46_1466626:
	.loc	19	126	0
	mul.lo.u64 	%rd20, %rd17, 16;
	add.u64 	%rd10, %rd18, %rd20;
	mov.f32 	%f122, 0f3f800000;   	// 1
	mov.f32 	%f123, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f124, %f122, %f123;
	mov.f32 	%f125, 0f437f0000;   	// 255
	mul.ftz.f32 	%f60, %f124, %f125;
	st.global.v4.f32 	[%rd10+0], {%f85,%f91,%f97,%f60};
$Lt_46_1466370:
	.loc	22	202	0
	@!%p5 bra 	$Lt_46_1467138;
	.loc	19	126	0
	mul.lo.u64 	%rd21, %rd17, 8;
	add.u64 	%rd8, %rd18, %rd21;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f114;
	mov.b32		%r62, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f111;
	mov.b32		%r63, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f117;
	mov.b32		%r64, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f60;
	mov.b32		%r65, %b1; }
	st.global.v4.u16 	[%rd8+8], {%r62,%r63,%r64,%r65};
	.loc	22	202	0
	bra.uni 	$Lt_46_1475586;
$Lt_46_1467138:
	.loc	19	126	0
	mul.lo.u64 	%rd22, %rd17, 16;
	add.u64 	%rd10, %rd18, %rd22;
	st.global.v4.f32 	[%rd10+16], {%f114,%f111,%f117,%f60};
	bra.uni 	$Lt_46_1475586;
$Lt_46_1450754:
	.loc	22	202	0
	setp.ge.s32 	%p6, %r12, %r13;
	@%p6 bra 	$Lt_46_1475586;
	.loc	20	469	0
	ld.param.u64 	%rd23, [__cudaparm_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inSrc];
	ld.param.s32 	%r66, [__cudaparm_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inSrcPitch];
	mul.lo.s32 	%r67, %r66, %r5;
	add.s32 	%r68, %r11, %r67;
	cvt.s64.s32 	%rd24, %r68;
	mul.wide.s32 	%rd25, %r68, 4;
	add.u64 	%rd26, %rd23, %rd25;
	ld.global.v4.u8 	{%r69,%r70,_,%r71}, [%rd26+0];
	cvt.rn.f32.u32 	%f126, %r71;
	cvt.rn.f32.u32 	%f127, %r70;
	cvt.rn.f32.u32 	%f128, %r69;
	.loc	20	490	0
	mov.f32 	%f129, 0f437f0000;   	// 255
	mov.f32 	%f130, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f131, %f129, %f130;
	ld.const.f32 	%f132, [kYCbCrOffset+4];
	mul.ftz.f32 	%f133, %f131, %f132;
	sub.ftz.f32 	%f134, %f127, %f133;
	ld.const.f32 	%f135, [kYCbCrOffset+0];
	mul.ftz.f32 	%f136, %f131, %f135;
	sub.ftz.f32 	%f137, %f128, %f136;
	ld.const.f32 	%f138, [kYCbCrOffset+8];
	mul.ftz.f32 	%f139, %f131, %f138;
	sub.ftz.f32 	%f140, %f126, %f139;
	ld.global.f32 	%f141, [k709YCbCr_To_RGB32f+16];
	mul.ftz.f32 	%f142, %f141, %f134;
	ld.global.f32 	%f143, [k709YCbCr_To_RGB32f+12];
	fma.rn.ftz.f32 	%f144, %f143, %f137, %f142;
	ld.global.f32 	%f145, [k709YCbCr_To_RGB32f+20];
	fma.rn.ftz.f32 	%f146, %f145, %f140, %f144;
	.loc	21	268	0
	ld.global.f32 	%f147, [k709YCbCr_To_RGB32f+4];
	mul.ftz.f32 	%f148, %f147, %f134;
	ld.global.f32 	%f149, [k709YCbCr_To_RGB32f+0];
	fma.rn.ftz.f32 	%f150, %f149, %f137, %f148;
	ld.global.f32 	%f151, [k709YCbCr_To_RGB32f+8];
	fma.rn.ftz.f32 	%f152, %f151, %f140, %f150;
	.loc	22	202	0
	ld.param.s32 	%r72, [__cudaparm_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDestPitch];
	mul.lo.s32 	%r73, %r72, %r5;
	ld.global.f32 	%f153, [k709YCbCr_To_RGB32f+28];
	mul.ftz.f32 	%f154, %f153, %f134;
	add.s32 	%r74, %r12, %r73;
	ld.global.f32 	%f155, [k709YCbCr_To_RGB32f+24];
	fma.rn.ftz.f32 	%f156, %f155, %f137, %f154;
	cvt.s64.s32 	%rd27, %r74;
	ld.global.f32 	%f157, [k709YCbCr_To_RGB32f+32];
	fma.rn.ftz.f32 	%f158, %f157, %f140, %f156;
	ld.param.u64 	%rd28, [__cudaparm_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDest];
	ld.param.s32 	%r75, [__cudaparm_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDestDeviceFormat];
	mov.u32 	%r76, 0;
	setp.ne.s32 	%p7, %r75, %r76;
	@%p7 bra 	$Lt_46_1475842;
	.loc	19	126	0
	mul.lo.u64 	%rd29, %rd27, 8;
	add.u64 	%rd8, %rd28, %rd29;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f158;
	mov.b32		%r77, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f146;
	mov.b32		%r78, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f152;
	mov.b32		%r79, %b1; }
	mov.f32 	%f159, 0f3f800000;   	// 1
	mov.f32 	%f160, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f161, %f159, %f160;
	mov.f32 	%f162, 0f437f0000;   	// 255
	mul.ftz.f32 	%f163, %f161, %f162;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f163;
	mov.b32		%r80, %b1; }
	st.global.v4.u16 	[%rd8+0], {%r77,%r78,%r79,%r80};
	.loc	22	202	0
	bra.uni 	$Lt_46_1475586;
$Lt_46_1475842:
	.loc	19	126	0
	mul.lo.u64 	%rd30, %rd27, 16;
	add.u64 	%rd10, %rd28, %rd30;
	mov.f32 	%f164, 0f3f800000;   	// 1
	mov.f32 	%f165, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f166, %f164, %f165;
	mov.f32 	%f167, 0f437f0000;   	// 255
	mul.ftz.f32 	%f168, %f166, %f167;
	st.global.v4.f32 	[%rd10+0], {%f158,%f146,%f152,%f168};
$Lt_46_1475586:
$Lt_46_1467394:
$Lt_46_1450498:
$Lt_46_1433602:
$Lt_46_1433090:
	.loc	22	202	0
	exit;
$LDWend_PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel:
	} // PixelFormatConvert_IR_PixelFormat_YUYV_422_8u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel

	.entry PixelFormatConvert_IR_PixelFormat_V210_422_10u_To_IR_PixelFormat_BGRA_4444_32f_Kernel (
		.param .u64 __cudaparm_PixelFormatConvert_IR_PixelFormat_V210_422_10u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inSrc,
		.param .s32 __cudaparm_PixelFormatConvert_IR_PixelFormat_V210_422_10u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inSrcPitch,
		.param .u32 __cudaparm_PixelFormatConvert_IR_PixelFormat_V210_422_10u_To_IR_PixelFormat_BGRA_4444_32f_Kernel___T2137,
		.param .u64 __cudaparm_PixelFormatConvert_IR_PixelFormat_V210_422_10u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDest,
		.param .s32 __cudaparm_PixelFormatConvert_IR_PixelFormat_V210_422_10u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDestPitch,
		.param .u32 __cudaparm_PixelFormatConvert_IR_PixelFormat_V210_422_10u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDestDeviceFormat,
		.param .s32 __cudaparm_PixelFormatConvert_IR_PixelFormat_V210_422_10u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inWidth,
		.param .s32 __cudaparm_PixelFormatConvert_IR_PixelFormat_V210_422_10u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inHeight)
	{
	.reg .u32 %r<101>;
	.reg .u64 %rd<22>;
	.reg .f32 %f<260>;
	.reg .pred %p<12>;
	// __cuda_local_var_304139_948_non_const_p1 = 16
	// __cuda_local_var_304139_952_non_const_p2 = 32
	// __cuda_local_var_304139_956_non_const_p3 = 48
	// __cuda_local_var_304139_960_non_const_p4 = 64
	// __cuda_local_var_304139_964_non_const_p5 = 80
	.loc	22	321	0
$LDWbegin_PixelFormatConvert_IR_PixelFormat_V210_422_10u_To_IR_PixelFormat_BGRA_4444_32f_Kernel:
	cvt.s32.u32 	%r1, %ctaid.x;
	cvt.s32.u32 	%r2, %ntid.x;
	mul.lo.s32 	%r3, %r1, %r2;
	cvt.s32.u32 	%r4, %ctaid.y;
	cvt.s32.u32 	%r5, %ntid.y;
	mul.lo.s32 	%r6, %r4, %r5;
	mov.u32 	%r7, %tid.x;
	add.u32 	%r8, %r3, %r7;
	mov.u32 	%r9, %tid.y;
	add.u32 	%r10, %r6, %r9;
	mul.lo.s32 	%r11, %r8, 6;
	ld.param.s32 	%r12, [__cudaparm_PixelFormatConvert_IR_PixelFormat_V210_422_10u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inWidth];
	ld.param.s32 	%r13, [__cudaparm_PixelFormatConvert_IR_PixelFormat_V210_422_10u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inHeight];
	set.gt.u32.s32 	%r14, %r13, %r10;
	neg.s32 	%r15, %r14;
	set.lt.u32.s32 	%r16, %r11, %r12;
	neg.s32 	%r17, %r16;
	and.b32 	%r18, %r15, %r17;
	mov.u32 	%r19, 0;
	setp.eq.s32 	%p1, %r18, %r19;
	@%p1 bra 	$Lt_47_2463490;
	.loc	19	115	0
	ld.param.u64 	%rd1, [__cudaparm_PixelFormatConvert_IR_PixelFormat_V210_422_10u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inSrc];
	ld.param.s32 	%r20, [__cudaparm_PixelFormatConvert_IR_PixelFormat_V210_422_10u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inSrcPitch];
	mul.lo.s32 	%r21, %r20, %r10;
	add.s32 	%r22, %r8, %r21;
	cvt.s64.s32 	%rd2, %r22;
	mul.wide.s32 	%rd3, %r22, 16;
	add.u64 	%rd4, %rd1, %rd3;
	ld.global.v4.u32 	{%r23,%r24,%r25,%r26}, [%rd4+0];
	.loc	22	321	0
	shl.b32 	%r27, %r25, 2;
	shr.u32 	%r28, %r27, 22;
	cvt.rn.f32.u32 	%f1, %r28;
	mov.f32 	%f2, %f1;
	shl.b32 	%r29, %r26, 12;
	shr.u32 	%r30, %r29, 22;
	cvt.rn.f32.u32 	%f3, %r30;
	mov.f32 	%f4, %f3;
	shl.b32 	%r31, %r23, 2;
	shr.u32 	%r32, %r31, 22;
	and.b32 	%r33, %r23, 1023;
	shl.b32 	%r34, %r23, 12;
	shr.u32 	%r35, %r34, 22;
	cvt.rn.f32.u32 	%f5, %r32;
	cvt.rn.f32.u32 	%f6, %r33;
	cvt.rn.f32.u32 	%f7, %r35;
	add.s32 	%r36, %r11, 1;
	setp.lt.s32 	%p2, %r36, %r12;
	ld.global.f32 	%f8, [k601YCbCr_To_RGB32f+20];
	ld.global.f32 	%f9, [k601YCbCr_To_RGB32f+12];
	ld.global.f32 	%f10, [k601YCbCr_To_RGB32f+16];
	ld.global.f32 	%f11, [k601YCbCr_To_RGB32f+8];
	ld.global.f32 	%f12, [k601YCbCr_To_RGB32f+0];
	ld.global.f32 	%f13, [k601YCbCr_To_RGB32f+4];
	ld.global.f32 	%f14, [k601YCbCr_To_RGB32f+32];
	ld.global.f32 	%f15, [k601YCbCr_To_RGB32f+24];
	ld.global.f32 	%f16, [k601YCbCr_To_RGB32f+28];
	ld.const.f32 	%f17, [kYCbCrOffset+8];
	ld.const.f32 	%f18, [kYCbCrOffset+0];
	ld.const.f32 	%f19, [kYCbCrOffset+4];
	@!%p2 bra 	$Lt_47_2464258;
	.loc	20	490	0
	mov.f32 	%f20, 0f447fc000;    	// 1023
	mov.f32 	%f21, 0f437f0000;    	// 255
	div.approx.ftz.f32 	%f22, %f20, %f21;
	mul.ftz.f32 	%f23, %f22, %f19;
	sub.ftz.f32 	%f24, %f6, %f23;
	mul.ftz.f32 	%f25, %f22, %f18;
	sub.ftz.f32 	%f26, %f7, %f25;
	mul.ftz.f32 	%f27, %f22, %f17;
	sub.ftz.f32 	%f28, %f5, %f27;
	mul.ftz.f32 	%f29, %f24, %f16;
	fma.rn.ftz.f32 	%f30, %f15, %f26, %f29;
	fma.rn.ftz.f32 	%f31, %f14, %f28, %f30;
	mul.ftz.f32 	%f32, %f24, %f13;
	fma.rn.ftz.f32 	%f33, %f12, %f26, %f32;
	fma.rn.ftz.f32 	%f34, %f11, %f28, %f33;
	mul.ftz.f32 	%f35, %f24, %f10;
	fma.rn.ftz.f32 	%f36, %f9, %f26, %f35;
	fma.rn.ftz.f32 	%f37, %f8, %f28, %f36;
	.loc	20	505	0
	mov.f32 	%f38, 0f437f0000;    	// 255
	mov.f32 	%f39, 0f447fc000;    	// 1023
	div.approx.ftz.f32 	%f40, %f38, %f39;
	mul.ftz.f32 	%f41, %f40, %f34;
	.loc	20	506	0
	mul.ftz.f32 	%f42, %f40, %f37;
	.loc	20	507	0
	mul.ftz.f32 	%f43, %f40, %f31;
	.loc	20	469	0
	and.b32 	%r37, %r25, 1023;
	cvt.rn.f32.u32 	%f44, %r37;
	add.ftz.f32 	%f45, %f44, %f5;
	cvt.rzi.ftz.u32.f32 	%r38, %f45;
	shr.u32 	%r39, %r38, 1;
	cvt.rn.f32.u32 	%f46, %r39;
	shl.b32 	%r40, %r24, 12;
	shr.u32 	%r41, %r40, 22;
	cvt.rn.f32.u32 	%f47, %r41;
	add.ftz.f32 	%f48, %f47, %f6;
	cvt.rzi.ftz.u32.f32 	%r42, %f48;
	shr.u32 	%r43, %r42, 1;
	cvt.rn.f32.u32 	%f49, %r43;
	and.b32 	%r44, %r24, 1023;
	cvt.rn.f32.u32 	%f50, %r44;
	.loc	20	490	0
	mul.ftz.f32 	%f51, %f22, %f19;
	sub.ftz.f32 	%f52, %f49, %f51;
	mul.ftz.f32 	%f53, %f22, %f18;
	sub.ftz.f32 	%f54, %f50, %f53;
	mul.ftz.f32 	%f55, %f22, %f17;
	sub.ftz.f32 	%f56, %f46, %f55;
	mul.ftz.f32 	%f57, %f52, %f16;
	fma.rn.ftz.f32 	%f58, %f15, %f54, %f57;
	fma.rn.ftz.f32 	%f59, %f14, %f56, %f58;
	mul.ftz.f32 	%f60, %f52, %f13;
	fma.rn.ftz.f32 	%f61, %f12, %f54, %f60;
	fma.rn.ftz.f32 	%f62, %f11, %f56, %f61;
	mul.ftz.f32 	%f63, %f52, %f10;
	fma.rn.ftz.f32 	%f64, %f9, %f54, %f63;
	fma.rn.ftz.f32 	%f65, %f8, %f56, %f64;
	.loc	20	505	0
	mul.ftz.f32 	%f66, %f40, %f62;
	.loc	20	506	0
	mul.ftz.f32 	%f67, %f40, %f65;
	.loc	20	507	0
	mul.ftz.f32 	%f68, %f59, %f40;
	.loc	22	78	0
	mov.f32 	%f69, %f68;
	mov.f32 	%f70, %f67;
	mov.f32 	%f71, %f66;
	mov.f32 	%f72, 0f3f800000;    	// 1
	mov.f32 	%f73, 0f447fc000;    	// 1023
	div.approx.ftz.f32 	%f74, %f72, %f73;
	mov.f32 	%f75, 0f447fc000;    	// 1023
	mul.ftz.f32 	%f76, %f74, %f75;
	mov.f32 	%f77, %f76;
	.loc	22	321	0
	shl.b32 	%r45, %r24, 2;
	shr.u32 	%r46, %r45, 22;
	cvt.rn.f32.u32 	%f78, %r46;
	add.s32 	%r47, %r11, 3;
	setp.le.s32 	%p3, %r12, %r47;
	@%p3 bra 	$Lt_47_2480130;
	.loc	20	490	0
	mul.ftz.f32 	%f79, %f22, %f19;
	sub.ftz.f32 	%f80, %f47, %f79;
	mul.ftz.f32 	%f81, %f22, %f18;
	sub.ftz.f32 	%f82, %f78, %f81;
	mul.ftz.f32 	%f83, %f22, %f17;
	sub.ftz.f32 	%f84, %f44, %f83;
	mul.ftz.f32 	%f85, %f80, %f16;
	fma.rn.ftz.f32 	%f86, %f15, %f82, %f85;
	fma.rn.ftz.f32 	%f87, %f14, %f84, %f86;
	mul.ftz.f32 	%f88, %f80, %f13;
	fma.rn.ftz.f32 	%f89, %f12, %f82, %f88;
	fma.rn.ftz.f32 	%f90, %f11, %f84, %f89;
	mul.ftz.f32 	%f91, %f80, %f10;
	fma.rn.ftz.f32 	%f92, %f9, %f82, %f91;
	fma.rn.ftz.f32 	%f93, %f8, %f84, %f92;
	.loc	20	505	0
	mul.ftz.f32 	%f94, %f90, %f40;
	.loc	20	506	0
	mul.ftz.f32 	%f95, %f93, %f40;
	.loc	20	507	0
	mul.ftz.f32 	%f96, %f87, %f40;
	.loc	22	60	0
	mov.f32 	%f97, %f96;
	mov.f32 	%f98, %f95;
	mov.f32 	%f99, %f94;
	.loc	20	469	0
	add.ftz.f32 	%f100, %f44, %f3;
	cvt.rzi.ftz.u32.f32 	%r48, %f100;
	shr.u32 	%r49, %r48, 1;
	cvt.rn.f32.u32 	%f101, %r49;
	add.ftz.f32 	%f102, %f47, %f1;
	cvt.rzi.ftz.u32.f32 	%r50, %f102;
	shr.u32 	%r51, %r50, 1;
	cvt.rn.f32.u32 	%f103, %r51;
	shl.b32 	%r52, %r25, 12;
	shr.u32 	%r53, %r52, 22;
	cvt.rn.f32.u32 	%f104, %r53;
	.loc	20	490	0
	mul.ftz.f32 	%f105, %f22, %f19;
	sub.ftz.f32 	%f106, %f103, %f105;
	mul.ftz.f32 	%f107, %f22, %f18;
	sub.ftz.f32 	%f108, %f104, %f107;
	mul.ftz.f32 	%f109, %f22, %f17;
	sub.ftz.f32 	%f110, %f101, %f109;
	mul.ftz.f32 	%f111, %f106, %f16;
	fma.rn.ftz.f32 	%f112, %f15, %f108, %f111;
	fma.rn.ftz.f32 	%f113, %f14, %f110, %f112;
	mul.ftz.f32 	%f114, %f106, %f13;
	fma.rn.ftz.f32 	%f115, %f12, %f108, %f114;
	fma.rn.ftz.f32 	%f116, %f11, %f110, %f115;
	mul.ftz.f32 	%f117, %f106, %f10;
	fma.rn.ftz.f32 	%f118, %f9, %f108, %f117;
	fma.rn.ftz.f32 	%f119, %f8, %f110, %f118;
	.loc	20	505	0
	mul.ftz.f32 	%f120, %f40, %f116;
	.loc	20	506	0
	mul.ftz.f32 	%f121, %f40, %f119;
	.loc	20	507	0
	mul.ftz.f32 	%f122, %f113, %f40;
	.loc	22	78	0
	mov.f32 	%f123, %f122;
	mov.f32 	%f124, %f121;
	mov.f32 	%f125, %f120;
	mov.f32 	%f126, %f76;
	.loc	22	321	0
	and.b32 	%r54, %r26, 1023;
	cvt.rn.f32.u32 	%f127, %r54;
	add.s32 	%r55, %r11, 5;
	setp.le.s32 	%p4, %r12, %r55;
	@%p4 bra 	$Lt_47_2496002;
	add.s32 	%r56, %r11, 6;
	setp.le.s32 	%p5, %r12, %r56;
	@%p5 bra 	$Lt_47_2496258;
	ld.global.u32 	%r57, [%rd4+16];
	and.b32 	%r58, %r57, 1023;
	cvt.rn.f32.u32 	%f2, %r58;
	shl.b32 	%r59, %r57, 2;
	shr.u32 	%r60, %r59, 22;
	cvt.rn.f32.u32 	%f4, %r60;
$Lt_47_2496258:
	.loc	20	490	0
	mul.ftz.f32 	%f128, %f22, %f19;
	sub.ftz.f32 	%f129, %f1, %f128;
	mul.ftz.f32 	%f130, %f22, %f18;
	sub.ftz.f32 	%f131, %f127, %f130;
	mul.ftz.f32 	%f132, %f22, %f17;
	sub.ftz.f32 	%f133, %f3, %f132;
	mul.ftz.f32 	%f134, %f129, %f16;
	fma.rn.ftz.f32 	%f135, %f15, %f131, %f134;
	fma.rn.ftz.f32 	%f136, %f14, %f133, %f135;
	mul.ftz.f32 	%f137, %f129, %f13;
	fma.rn.ftz.f32 	%f138, %f12, %f131, %f137;
	fma.rn.ftz.f32 	%f139, %f11, %f133, %f138;
	mul.ftz.f32 	%f140, %f129, %f10;
	fma.rn.ftz.f32 	%f141, %f9, %f131, %f140;
	fma.rn.ftz.f32 	%f142, %f8, %f133, %f141;
	.loc	20	505	0
	mul.ftz.f32 	%f143, %f139, %f40;
	.loc	20	506	0
	mul.ftz.f32 	%f144, %f142, %f40;
	.loc	20	507	0
	mul.ftz.f32 	%f145, %f136, %f40;
	.loc	22	60	0
	mov.f32 	%f146, %f145;
	mov.f32 	%f147, %f144;
	mov.f32 	%f148, %f143;
	.loc	20	469	0
	add.ftz.f32 	%f149, %f3, %f4;
	cvt.rzi.ftz.u32.f32 	%r61, %f149;
	shr.u32 	%r62, %r61, 1;
	cvt.rn.f32.u32 	%f150, %r62;
	add.ftz.f32 	%f151, %f1, %f2;
	cvt.rzi.ftz.u32.f32 	%r63, %f151;
	shr.u32 	%r64, %r63, 1;
	cvt.rn.f32.u32 	%f152, %r64;
	shl.b32 	%r65, %r26, 2;
	shr.u32 	%r66, %r65, 22;
	cvt.rn.f32.u32 	%f153, %r66;
	.loc	20	490	0
	mul.ftz.f32 	%f154, %f22, %f19;
	sub.ftz.f32 	%f155, %f152, %f154;
	mul.ftz.f32 	%f156, %f22, %f18;
	sub.ftz.f32 	%f157, %f153, %f156;
	mul.ftz.f32 	%f158, %f22, %f17;
	sub.ftz.f32 	%f159, %f150, %f158;
	mul.ftz.f32 	%f160, %f155, %f16;
	fma.rn.ftz.f32 	%f161, %f15, %f157, %f160;
	fma.rn.ftz.f32 	%f162, %f14, %f159, %f161;
	mul.ftz.f32 	%f163, %f155, %f13;
	fma.rn.ftz.f32 	%f164, %f12, %f157, %f163;
	fma.rn.ftz.f32 	%f165, %f11, %f159, %f164;
	mul.ftz.f32 	%f166, %f155, %f10;
	fma.rn.ftz.f32 	%f167, %f9, %f157, %f166;
	fma.rn.ftz.f32 	%f168, %f8, %f159, %f167;
	.loc	20	505	0
	mul.ftz.f32 	%f169, %f40, %f165;
	.loc	20	506	0
	mul.ftz.f32 	%f170, %f40, %f168;
	.loc	20	507	0
	mul.ftz.f32 	%f171, %f162, %f40;
	.loc	22	78	0
	mov.f32 	%f172, %f171;
	mov.f32 	%f173, %f170;
	mov.f32 	%f174, %f169;
	mov.f32 	%f175, %f76;
	.loc	22	321	0
	bra.uni 	$Lt_47_2495746;
$Lt_47_2496002:
	.loc	20	490	0
	mul.ftz.f32 	%f176, %f22, %f19;
	sub.ftz.f32 	%f177, %f1, %f176;
	mul.ftz.f32 	%f178, %f22, %f18;
	sub.ftz.f32 	%f179, %f127, %f178;
	mul.ftz.f32 	%f180, %f22, %f17;
	sub.ftz.f32 	%f181, %f3, %f180;
	mul.ftz.f32 	%f182, %f177, %f16;
	fma.rn.ftz.f32 	%f183, %f15, %f179, %f182;
	fma.rn.ftz.f32 	%f184, %f14, %f181, %f183;
	mul.ftz.f32 	%f185, %f177, %f13;
	fma.rn.ftz.f32 	%f186, %f12, %f179, %f185;
	fma.rn.ftz.f32 	%f187, %f11, %f181, %f186;
	mul.ftz.f32 	%f188, %f177, %f10;
	fma.rn.ftz.f32 	%f189, %f9, %f179, %f188;
	fma.rn.ftz.f32 	%f190, %f8, %f181, %f189;
	.loc	20	505	0
	mul.ftz.f32 	%f191, %f187, %f40;
	.loc	20	506	0
	mul.ftz.f32 	%f192, %f190, %f40;
	.loc	20	507	0
	mul.ftz.f32 	%f193, %f184, %f40;
	.loc	22	321	0
	mov.f32 	%f146, %f193;
	mov.f32 	%f147, %f192;
	mov.f32 	%f148, %f191;
$Lt_47_2495746:
	mov.f32 	%f194, %f76;
	bra.uni 	$Lt_47_2479874;
$Lt_47_2480130:
	.loc	20	490	0
	mul.ftz.f32 	%f195, %f22, %f19;
	sub.ftz.f32 	%f196, %f47, %f195;
	mul.ftz.f32 	%f197, %f22, %f18;
	sub.ftz.f32 	%f198, %f78, %f197;
	mul.ftz.f32 	%f199, %f22, %f17;
	sub.ftz.f32 	%f200, %f44, %f199;
	mul.ftz.f32 	%f201, %f196, %f16;
	fma.rn.ftz.f32 	%f202, %f15, %f198, %f201;
	fma.rn.ftz.f32 	%f203, %f14, %f200, %f202;
	mul.ftz.f32 	%f204, %f196, %f13;
	fma.rn.ftz.f32 	%f205, %f12, %f198, %f204;
	fma.rn.ftz.f32 	%f206, %f11, %f200, %f205;
	mul.ftz.f32 	%f207, %f196, %f10;
	fma.rn.ftz.f32 	%f208, %f9, %f198, %f207;
	fma.rn.ftz.f32 	%f209, %f8, %f200, %f208;
	.loc	20	505	0
	mul.ftz.f32 	%f210, %f206, %f40;
	.loc	20	506	0
	mul.ftz.f32 	%f211, %f209, %f40;
	.loc	20	507	0
	mul.ftz.f32 	%f212, %f203, %f40;
	.loc	22	321	0
	mov.f32 	%f97, %f212;
	mov.f32 	%f98, %f211;
	mov.f32 	%f99, %f210;
$Lt_47_2479874:
	mov.f32 	%f213, %f76;
	bra.uni 	$Lt_47_2464002;
$Lt_47_2464258:
	.loc	20	490	0
	mov.f32 	%f214, 0f447fc000;   	// 1023
	mov.f32 	%f215, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f216, %f214, %f215;
	mul.ftz.f32 	%f217, %f216, %f19;
	sub.ftz.f32 	%f218, %f6, %f217;
	mul.ftz.f32 	%f219, %f216, %f18;
	sub.ftz.f32 	%f220, %f7, %f219;
	mul.ftz.f32 	%f221, %f216, %f17;
	sub.ftz.f32 	%f222, %f5, %f221;
	mul.ftz.f32 	%f223, %f218, %f16;
	fma.rn.ftz.f32 	%f224, %f15, %f220, %f223;
	fma.rn.ftz.f32 	%f225, %f14, %f222, %f224;
	mul.ftz.f32 	%f226, %f218, %f13;
	fma.rn.ftz.f32 	%f227, %f12, %f220, %f226;
	fma.rn.ftz.f32 	%f228, %f11, %f222, %f227;
	mul.ftz.f32 	%f229, %f218, %f10;
	fma.rn.ftz.f32 	%f230, %f9, %f220, %f229;
	fma.rn.ftz.f32 	%f231, %f8, %f222, %f230;
	.loc	20	505	0
	mov.f32 	%f232, 0f437f0000;   	// 255
	mov.f32 	%f233, 0f447fc000;   	// 1023
	div.approx.ftz.f32 	%f234, %f232, %f233;
	mul.ftz.f32 	%f41, %f234, %f228;
	.loc	20	506	0
	mul.ftz.f32 	%f42, %f234, %f231;
	.loc	20	507	0
	mul.ftz.f32 	%f43, %f234, %f225;
	.loc	22	321	0
	mov.f32 	%f235, 0f3f800000;   	// 1
	mov.f32 	%f236, 0f447fc000;   	// 1023
	div.approx.ftz.f32 	%f237, %f235, %f236;
	mov.f32 	%f238, 0f447fc000;   	// 1023
	mul.ftz.f32 	%f76, %f237, %f238;
$Lt_47_2464002:
	ld.param.s32 	%r67, [__cudaparm_PixelFormatConvert_IR_PixelFormat_V210_422_10u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDestDeviceFormat];
	mov.s32 	%r68, 0;
	setp.eq.s32 	%p6, %r67, %r68;
	ld.param.s32 	%r69, [__cudaparm_PixelFormatConvert_IR_PixelFormat_V210_422_10u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDestPitch];
	mul.lo.s32 	%r70, %r69, %r10;
	add.s32 	%r71, %r11, %r70;
	cvt.s64.s32 	%rd5, %r71;
	ld.param.u64 	%rd6, [__cudaparm_PixelFormatConvert_IR_PixelFormat_V210_422_10u_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDest];
	@!%p6 bra 	$Lt_47_2535426;
	.loc	19	126	0
	mul.lo.u64 	%rd7, %rd5, 8;
	add.u64 	%rd8, %rd6, %rd7;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f43;
	mov.b32		%r72, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f42;
	mov.b32		%r73, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f41;
	mov.b32		%r74, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f76;
	mov.b32		%r75, %b1; }
	st.global.v4.u16 	[%rd8+0], {%r72,%r73,%r74,%r75};
	.loc	22	321	0
	bra.uni 	$Lt_47_2535170;
$Lt_47_2535426:
	.loc	19	126	0
	mul.lo.u64 	%rd9, %rd5, 16;
	add.u64 	%rd10, %rd6, %rd9;
	st.global.v4.f32 	[%rd10+0], {%f43,%f42,%f41,%f76};
$Lt_47_2535170:
	.loc	22	321	0
	@!%p2 bra 	$Lt_47_2535682;
	mov.f32 	%f239, %f77;
	mov.f32 	%f240, %f71;
	mov.f32 	%f241, %f70;
	mov.f32 	%f242, %f69;
	@!%p6 bra 	$Lt_47_2536450;
	.loc	19	126	0
	mul.lo.u64 	%rd11, %rd5, 8;
	add.u64 	%rd8, %rd6, %rd11;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f242;
	mov.b32		%r76, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f241;
	mov.b32		%r77, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f240;
	mov.b32		%r78, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f239;
	mov.b32		%r79, %b1; }
	st.global.v4.u16 	[%rd8+8], {%r76,%r77,%r78,%r79};
	.loc	22	321	0
	bra.uni 	$Lt_47_2536194;
$Lt_47_2536450:
	.loc	19	126	0
	mul.lo.u64 	%rd12, %rd5, 16;
	add.u64 	%rd10, %rd6, %rd12;
	st.global.v4.f32 	[%rd10+16], {%f242,%f241,%f240,%f239};
$Lt_47_2536194:
	.loc	22	321	0
	add.s32 	%r80, %r11, 2;
	setp.le.s32 	%p7, %r12, %r80;
	@%p7 bra 	$Lt_47_2536706;
	mov.f32 	%f243, %f213;
	mov.f32 	%f244, %f99;
	mov.f32 	%f245, %f98;
	mov.f32 	%f246, %f97;
	@!%p6 bra 	$Lt_47_2537474;
	.loc	19	126	0
	mul.lo.u64 	%rd13, %rd5, 8;
	add.u64 	%rd8, %rd6, %rd13;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f246;
	mov.b32		%r81, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f245;
	mov.b32		%r82, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f244;
	mov.b32		%r83, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f243;
	mov.b32		%r84, %b1; }
	st.global.v4.u16 	[%rd8+16], {%r81,%r82,%r83,%r84};
	.loc	22	321	0
	bra.uni 	$Lt_47_2537218;
$Lt_47_2537474:
	.loc	19	126	0
	mul.lo.u64 	%rd14, %rd5, 16;
	add.u64 	%rd10, %rd6, %rd14;
	st.global.v4.f32 	[%rd10+32], {%f246,%f245,%f244,%f243};
$Lt_47_2537218:
	.loc	22	321	0
	add.s32 	%r85, %r11, 3;
	setp.le.s32 	%p8, %r12, %r85;
	@%p8 bra 	$Lt_47_2537730;
	mov.f32 	%f247, %f126;
	mov.f32 	%f248, %f125;
	mov.f32 	%f249, %f124;
	mov.f32 	%f250, %f123;
	@!%p6 bra 	$Lt_47_2538498;
	.loc	19	126	0
	mul.lo.u64 	%rd15, %rd5, 8;
	add.u64 	%rd8, %rd6, %rd15;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f250;
	mov.b32		%r86, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f249;
	mov.b32		%r87, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f248;
	mov.b32		%r88, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f247;
	mov.b32		%r89, %b1; }
	st.global.v4.u16 	[%rd8+24], {%r86,%r87,%r88,%r89};
	.loc	22	321	0
	bra.uni 	$Lt_47_2538242;
$Lt_47_2538498:
	.loc	19	126	0
	mul.lo.u64 	%rd16, %rd5, 16;
	add.u64 	%rd10, %rd6, %rd16;
	st.global.v4.f32 	[%rd10+48], {%f250,%f249,%f248,%f247};
$Lt_47_2538242:
	.loc	22	321	0
	add.s32 	%r90, %r11, 4;
	setp.le.s32 	%p9, %r12, %r90;
	@%p9 bra 	$Lt_47_2538754;
	mov.f32 	%f251, %f194;
	mov.f32 	%f252, %f148;
	mov.f32 	%f253, %f147;
	mov.f32 	%f254, %f146;
	@!%p6 bra 	$Lt_47_2539522;
	.loc	19	126	0
	mul.lo.u64 	%rd17, %rd5, 8;
	add.u64 	%rd8, %rd6, %rd17;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f254;
	mov.b32		%r91, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f253;
	mov.b32		%r92, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f252;
	mov.b32		%r93, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f251;
	mov.b32		%r94, %b1; }
	st.global.v4.u16 	[%rd8+32], {%r91,%r92,%r93,%r94};
	.loc	22	321	0
	bra.uni 	$Lt_47_2539266;
$Lt_47_2539522:
	.loc	19	126	0
	mul.lo.u64 	%rd18, %rd5, 16;
	add.u64 	%rd10, %rd6, %rd18;
	st.global.v4.f32 	[%rd10+64], {%f254,%f253,%f252,%f251};
$Lt_47_2539266:
	.loc	22	321	0
	add.s32 	%r95, %r11, 5;
	setp.le.s32 	%p10, %r12, %r95;
	@%p10 bra 	$Lt_47_2539778;
	mov.f32 	%f255, %f175;
	mov.f32 	%f256, %f174;
	mov.f32 	%f257, %f173;
	mov.f32 	%f258, %f172;
	@!%p6 bra 	$Lt_47_2540546;
	.loc	19	126	0
	mul.lo.u64 	%rd19, %rd5, 8;
	add.u64 	%rd8, %rd6, %rd19;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f258;
	mov.b32		%r96, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f257;
	mov.b32		%r97, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f256;
	mov.b32		%r98, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f255;
	mov.b32		%r99, %b1; }
	st.global.v4.u16 	[%rd8+40], {%r96,%r97,%r98,%r99};
	.loc	22	321	0
	bra.uni 	$Lt_47_2540290;
$Lt_47_2540546:
	.loc	19	126	0
	mul.lo.u64 	%rd20, %rd5, 16;
	add.u64 	%rd10, %rd6, %rd20;
	st.global.v4.f32 	[%rd10+80], {%f258,%f257,%f256,%f255};
$Lt_47_2540290:
$Lt_47_2539778:
$Lt_47_2538754:
$Lt_47_2537730:
$Lt_47_2536706:
$Lt_47_2535682:
$Lt_47_2463490:
	.loc	22	321	0
	exit;
$LDWend_PixelFormatConvert_IR_PixelFormat_V210_422_10u_To_IR_PixelFormat_BGRA_4444_32f_Kernel:
	} // PixelFormatConvert_IR_PixelFormat_V210_422_10u_To_IR_PixelFormat_BGRA_4444_32f_Kernel

	.entry PixelFormatConvert_IR_PixelFormat_V210_422_10u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel (
		.param .u64 __cudaparm_PixelFormatConvert_IR_PixelFormat_V210_422_10u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inSrc,
		.param .s32 __cudaparm_PixelFormatConvert_IR_PixelFormat_V210_422_10u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inSrcPitch,
		.param .u32 __cudaparm_PixelFormatConvert_IR_PixelFormat_V210_422_10u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel___T2168,
		.param .u64 __cudaparm_PixelFormatConvert_IR_PixelFormat_V210_422_10u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDest,
		.param .s32 __cudaparm_PixelFormatConvert_IR_PixelFormat_V210_422_10u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDestPitch,
		.param .u32 __cudaparm_PixelFormatConvert_IR_PixelFormat_V210_422_10u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDestDeviceFormat,
		.param .s32 __cudaparm_PixelFormatConvert_IR_PixelFormat_V210_422_10u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inWidth,
		.param .s32 __cudaparm_PixelFormatConvert_IR_PixelFormat_V210_422_10u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inHeight)
	{
	.reg .u32 %r<101>;
	.reg .u64 %rd<22>;
	.reg .f32 %f<260>;
	.reg .pred %p<12>;
	// __cuda_local_var_304140_952_non_const_p1 = 16
	// __cuda_local_var_304140_956_non_const_p2 = 32
	// __cuda_local_var_304140_960_non_const_p3 = 48
	// __cuda_local_var_304140_964_non_const_p4 = 64
	// __cuda_local_var_304140_968_non_const_p5 = 80
	.loc	22	322	0
$LDWbegin_PixelFormatConvert_IR_PixelFormat_V210_422_10u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel:
	cvt.s32.u32 	%r1, %ctaid.x;
	cvt.s32.u32 	%r2, %ntid.x;
	mul.lo.s32 	%r3, %r1, %r2;
	cvt.s32.u32 	%r4, %ctaid.y;
	cvt.s32.u32 	%r5, %ntid.y;
	mul.lo.s32 	%r6, %r4, %r5;
	mov.u32 	%r7, %tid.x;
	add.u32 	%r8, %r3, %r7;
	mov.u32 	%r9, %tid.y;
	add.u32 	%r10, %r6, %r9;
	mul.lo.s32 	%r11, %r8, 6;
	ld.param.s32 	%r12, [__cudaparm_PixelFormatConvert_IR_PixelFormat_V210_422_10u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inWidth];
	ld.param.s32 	%r13, [__cudaparm_PixelFormatConvert_IR_PixelFormat_V210_422_10u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inHeight];
	set.gt.u32.s32 	%r14, %r13, %r10;
	neg.s32 	%r15, %r14;
	set.lt.u32.s32 	%r16, %r11, %r12;
	neg.s32 	%r17, %r16;
	and.b32 	%r18, %r15, %r17;
	mov.u32 	%r19, 0;
	setp.eq.s32 	%p1, %r18, %r19;
	@%p1 bra 	$Lt_48_2463490;
	.loc	19	115	0
	ld.param.u64 	%rd1, [__cudaparm_PixelFormatConvert_IR_PixelFormat_V210_422_10u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inSrc];
	ld.param.s32 	%r20, [__cudaparm_PixelFormatConvert_IR_PixelFormat_V210_422_10u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inSrcPitch];
	mul.lo.s32 	%r21, %r20, %r10;
	add.s32 	%r22, %r8, %r21;
	cvt.s64.s32 	%rd2, %r22;
	mul.wide.s32 	%rd3, %r22, 16;
	add.u64 	%rd4, %rd1, %rd3;
	ld.global.v4.u32 	{%r23,%r24,%r25,%r26}, [%rd4+0];
	.loc	22	322	0
	shl.b32 	%r27, %r25, 2;
	shr.u32 	%r28, %r27, 22;
	cvt.rn.f32.u32 	%f1, %r28;
	mov.f32 	%f2, %f1;
	shl.b32 	%r29, %r26, 12;
	shr.u32 	%r30, %r29, 22;
	cvt.rn.f32.u32 	%f3, %r30;
	mov.f32 	%f4, %f3;
	shl.b32 	%r31, %r23, 2;
	shr.u32 	%r32, %r31, 22;
	and.b32 	%r33, %r23, 1023;
	shl.b32 	%r34, %r23, 12;
	shr.u32 	%r35, %r34, 22;
	cvt.rn.f32.u32 	%f5, %r32;
	cvt.rn.f32.u32 	%f6, %r33;
	cvt.rn.f32.u32 	%f7, %r35;
	add.s32 	%r36, %r11, 1;
	setp.lt.s32 	%p2, %r36, %r12;
	ld.global.f32 	%f8, [k709YCbCr_To_RGB32f+20];
	ld.global.f32 	%f9, [k709YCbCr_To_RGB32f+12];
	ld.global.f32 	%f10, [k709YCbCr_To_RGB32f+16];
	ld.global.f32 	%f11, [k709YCbCr_To_RGB32f+8];
	ld.global.f32 	%f12, [k709YCbCr_To_RGB32f+0];
	ld.global.f32 	%f13, [k709YCbCr_To_RGB32f+4];
	ld.global.f32 	%f14, [k709YCbCr_To_RGB32f+32];
	ld.global.f32 	%f15, [k709YCbCr_To_RGB32f+24];
	ld.global.f32 	%f16, [k709YCbCr_To_RGB32f+28];
	ld.const.f32 	%f17, [kYCbCrOffset+8];
	ld.const.f32 	%f18, [kYCbCrOffset+0];
	ld.const.f32 	%f19, [kYCbCrOffset+4];
	@!%p2 bra 	$Lt_48_2464258;
	.loc	20	490	0
	mov.f32 	%f20, 0f447fc000;    	// 1023
	mov.f32 	%f21, 0f437f0000;    	// 255
	div.approx.ftz.f32 	%f22, %f20, %f21;
	mul.ftz.f32 	%f23, %f22, %f19;
	sub.ftz.f32 	%f24, %f6, %f23;
	mul.ftz.f32 	%f25, %f22, %f18;
	sub.ftz.f32 	%f26, %f7, %f25;
	mul.ftz.f32 	%f27, %f22, %f17;
	sub.ftz.f32 	%f28, %f5, %f27;
	mul.ftz.f32 	%f29, %f24, %f16;
	fma.rn.ftz.f32 	%f30, %f15, %f26, %f29;
	fma.rn.ftz.f32 	%f31, %f14, %f28, %f30;
	mul.ftz.f32 	%f32, %f24, %f13;
	fma.rn.ftz.f32 	%f33, %f12, %f26, %f32;
	fma.rn.ftz.f32 	%f34, %f11, %f28, %f33;
	mul.ftz.f32 	%f35, %f24, %f10;
	fma.rn.ftz.f32 	%f36, %f9, %f26, %f35;
	fma.rn.ftz.f32 	%f37, %f8, %f28, %f36;
	.loc	20	505	0
	mov.f32 	%f38, 0f437f0000;    	// 255
	mov.f32 	%f39, 0f447fc000;    	// 1023
	div.approx.ftz.f32 	%f40, %f38, %f39;
	mul.ftz.f32 	%f41, %f40, %f34;
	.loc	20	506	0
	mul.ftz.f32 	%f42, %f40, %f37;
	.loc	20	507	0
	mul.ftz.f32 	%f43, %f40, %f31;
	.loc	20	469	0
	and.b32 	%r37, %r25, 1023;
	cvt.rn.f32.u32 	%f44, %r37;
	add.ftz.f32 	%f45, %f44, %f5;
	cvt.rzi.ftz.u32.f32 	%r38, %f45;
	shr.u32 	%r39, %r38, 1;
	cvt.rn.f32.u32 	%f46, %r39;
	shl.b32 	%r40, %r24, 12;
	shr.u32 	%r41, %r40, 22;
	cvt.rn.f32.u32 	%f47, %r41;
	add.ftz.f32 	%f48, %f47, %f6;
	cvt.rzi.ftz.u32.f32 	%r42, %f48;
	shr.u32 	%r43, %r42, 1;
	cvt.rn.f32.u32 	%f49, %r43;
	and.b32 	%r44, %r24, 1023;
	cvt.rn.f32.u32 	%f50, %r44;
	.loc	20	490	0
	mul.ftz.f32 	%f51, %f22, %f19;
	sub.ftz.f32 	%f52, %f49, %f51;
	mul.ftz.f32 	%f53, %f22, %f18;
	sub.ftz.f32 	%f54, %f50, %f53;
	mul.ftz.f32 	%f55, %f22, %f17;
	sub.ftz.f32 	%f56, %f46, %f55;
	mul.ftz.f32 	%f57, %f52, %f16;
	fma.rn.ftz.f32 	%f58, %f15, %f54, %f57;
	fma.rn.ftz.f32 	%f59, %f14, %f56, %f58;
	mul.ftz.f32 	%f60, %f52, %f13;
	fma.rn.ftz.f32 	%f61, %f12, %f54, %f60;
	fma.rn.ftz.f32 	%f62, %f11, %f56, %f61;
	mul.ftz.f32 	%f63, %f52, %f10;
	fma.rn.ftz.f32 	%f64, %f9, %f54, %f63;
	fma.rn.ftz.f32 	%f65, %f8, %f56, %f64;
	.loc	20	505	0
	mul.ftz.f32 	%f66, %f40, %f62;
	.loc	20	506	0
	mul.ftz.f32 	%f67, %f40, %f65;
	.loc	20	507	0
	mul.ftz.f32 	%f68, %f59, %f40;
	.loc	22	78	0
	mov.f32 	%f69, %f68;
	mov.f32 	%f70, %f67;
	mov.f32 	%f71, %f66;
	mov.f32 	%f72, 0f3f800000;    	// 1
	mov.f32 	%f73, 0f447fc000;    	// 1023
	div.approx.ftz.f32 	%f74, %f72, %f73;
	mov.f32 	%f75, 0f447fc000;    	// 1023
	mul.ftz.f32 	%f76, %f74, %f75;
	mov.f32 	%f77, %f76;
	.loc	22	322	0
	shl.b32 	%r45, %r24, 2;
	shr.u32 	%r46, %r45, 22;
	cvt.rn.f32.u32 	%f78, %r46;
	add.s32 	%r47, %r11, 3;
	setp.le.s32 	%p3, %r12, %r47;
	@%p3 bra 	$Lt_48_2480130;
	.loc	20	490	0
	mul.ftz.f32 	%f79, %f22, %f19;
	sub.ftz.f32 	%f80, %f47, %f79;
	mul.ftz.f32 	%f81, %f22, %f18;
	sub.ftz.f32 	%f82, %f78, %f81;
	mul.ftz.f32 	%f83, %f22, %f17;
	sub.ftz.f32 	%f84, %f44, %f83;
	mul.ftz.f32 	%f85, %f80, %f16;
	fma.rn.ftz.f32 	%f86, %f15, %f82, %f85;
	fma.rn.ftz.f32 	%f87, %f14, %f84, %f86;
	mul.ftz.f32 	%f88, %f80, %f13;
	fma.rn.ftz.f32 	%f89, %f12, %f82, %f88;
	fma.rn.ftz.f32 	%f90, %f11, %f84, %f89;
	mul.ftz.f32 	%f91, %f80, %f10;
	fma.rn.ftz.f32 	%f92, %f9, %f82, %f91;
	fma.rn.ftz.f32 	%f93, %f8, %f84, %f92;
	.loc	20	505	0
	mul.ftz.f32 	%f94, %f90, %f40;
	.loc	20	506	0
	mul.ftz.f32 	%f95, %f93, %f40;
	.loc	20	507	0
	mul.ftz.f32 	%f96, %f87, %f40;
	.loc	22	60	0
	mov.f32 	%f97, %f96;
	mov.f32 	%f98, %f95;
	mov.f32 	%f99, %f94;
	.loc	20	469	0
	add.ftz.f32 	%f100, %f44, %f3;
	cvt.rzi.ftz.u32.f32 	%r48, %f100;
	shr.u32 	%r49, %r48, 1;
	cvt.rn.f32.u32 	%f101, %r49;
	add.ftz.f32 	%f102, %f47, %f1;
	cvt.rzi.ftz.u32.f32 	%r50, %f102;
	shr.u32 	%r51, %r50, 1;
	cvt.rn.f32.u32 	%f103, %r51;
	shl.b32 	%r52, %r25, 12;
	shr.u32 	%r53, %r52, 22;
	cvt.rn.f32.u32 	%f104, %r53;
	.loc	20	490	0
	mul.ftz.f32 	%f105, %f22, %f19;
	sub.ftz.f32 	%f106, %f103, %f105;
	mul.ftz.f32 	%f107, %f22, %f18;
	sub.ftz.f32 	%f108, %f104, %f107;
	mul.ftz.f32 	%f109, %f22, %f17;
	sub.ftz.f32 	%f110, %f101, %f109;
	mul.ftz.f32 	%f111, %f106, %f16;
	fma.rn.ftz.f32 	%f112, %f15, %f108, %f111;
	fma.rn.ftz.f32 	%f113, %f14, %f110, %f112;
	mul.ftz.f32 	%f114, %f106, %f13;
	fma.rn.ftz.f32 	%f115, %f12, %f108, %f114;
	fma.rn.ftz.f32 	%f116, %f11, %f110, %f115;
	mul.ftz.f32 	%f117, %f106, %f10;
	fma.rn.ftz.f32 	%f118, %f9, %f108, %f117;
	fma.rn.ftz.f32 	%f119, %f8, %f110, %f118;
	.loc	20	505	0
	mul.ftz.f32 	%f120, %f40, %f116;
	.loc	20	506	0
	mul.ftz.f32 	%f121, %f40, %f119;
	.loc	20	507	0
	mul.ftz.f32 	%f122, %f113, %f40;
	.loc	22	78	0
	mov.f32 	%f123, %f122;
	mov.f32 	%f124, %f121;
	mov.f32 	%f125, %f120;
	mov.f32 	%f126, %f76;
	.loc	22	322	0
	and.b32 	%r54, %r26, 1023;
	cvt.rn.f32.u32 	%f127, %r54;
	add.s32 	%r55, %r11, 5;
	setp.le.s32 	%p4, %r12, %r55;
	@%p4 bra 	$Lt_48_2496002;
	add.s32 	%r56, %r11, 6;
	setp.le.s32 	%p5, %r12, %r56;
	@%p5 bra 	$Lt_48_2496258;
	ld.global.u32 	%r57, [%rd4+16];
	and.b32 	%r58, %r57, 1023;
	cvt.rn.f32.u32 	%f2, %r58;
	shl.b32 	%r59, %r57, 2;
	shr.u32 	%r60, %r59, 22;
	cvt.rn.f32.u32 	%f4, %r60;
$Lt_48_2496258:
	.loc	20	490	0
	mul.ftz.f32 	%f128, %f22, %f19;
	sub.ftz.f32 	%f129, %f1, %f128;
	mul.ftz.f32 	%f130, %f22, %f18;
	sub.ftz.f32 	%f131, %f127, %f130;
	mul.ftz.f32 	%f132, %f22, %f17;
	sub.ftz.f32 	%f133, %f3, %f132;
	mul.ftz.f32 	%f134, %f129, %f16;
	fma.rn.ftz.f32 	%f135, %f15, %f131, %f134;
	fma.rn.ftz.f32 	%f136, %f14, %f133, %f135;
	mul.ftz.f32 	%f137, %f129, %f13;
	fma.rn.ftz.f32 	%f138, %f12, %f131, %f137;
	fma.rn.ftz.f32 	%f139, %f11, %f133, %f138;
	mul.ftz.f32 	%f140, %f129, %f10;
	fma.rn.ftz.f32 	%f141, %f9, %f131, %f140;
	fma.rn.ftz.f32 	%f142, %f8, %f133, %f141;
	.loc	20	505	0
	mul.ftz.f32 	%f143, %f139, %f40;
	.loc	20	506	0
	mul.ftz.f32 	%f144, %f142, %f40;
	.loc	20	507	0
	mul.ftz.f32 	%f145, %f136, %f40;
	.loc	22	60	0
	mov.f32 	%f146, %f145;
	mov.f32 	%f147, %f144;
	mov.f32 	%f148, %f143;
	.loc	20	469	0
	add.ftz.f32 	%f149, %f3, %f4;
	cvt.rzi.ftz.u32.f32 	%r61, %f149;
	shr.u32 	%r62, %r61, 1;
	cvt.rn.f32.u32 	%f150, %r62;
	add.ftz.f32 	%f151, %f1, %f2;
	cvt.rzi.ftz.u32.f32 	%r63, %f151;
	shr.u32 	%r64, %r63, 1;
	cvt.rn.f32.u32 	%f152, %r64;
	shl.b32 	%r65, %r26, 2;
	shr.u32 	%r66, %r65, 22;
	cvt.rn.f32.u32 	%f153, %r66;
	.loc	20	490	0
	mul.ftz.f32 	%f154, %f22, %f19;
	sub.ftz.f32 	%f155, %f152, %f154;
	mul.ftz.f32 	%f156, %f22, %f18;
	sub.ftz.f32 	%f157, %f153, %f156;
	mul.ftz.f32 	%f158, %f22, %f17;
	sub.ftz.f32 	%f159, %f150, %f158;
	mul.ftz.f32 	%f160, %f155, %f16;
	fma.rn.ftz.f32 	%f161, %f15, %f157, %f160;
	fma.rn.ftz.f32 	%f162, %f14, %f159, %f161;
	mul.ftz.f32 	%f163, %f155, %f13;
	fma.rn.ftz.f32 	%f164, %f12, %f157, %f163;
	fma.rn.ftz.f32 	%f165, %f11, %f159, %f164;
	mul.ftz.f32 	%f166, %f155, %f10;
	fma.rn.ftz.f32 	%f167, %f9, %f157, %f166;
	fma.rn.ftz.f32 	%f168, %f8, %f159, %f167;
	.loc	20	505	0
	mul.ftz.f32 	%f169, %f40, %f165;
	.loc	20	506	0
	mul.ftz.f32 	%f170, %f40, %f168;
	.loc	20	507	0
	mul.ftz.f32 	%f171, %f162, %f40;
	.loc	22	78	0
	mov.f32 	%f172, %f171;
	mov.f32 	%f173, %f170;
	mov.f32 	%f174, %f169;
	mov.f32 	%f175, %f76;
	.loc	22	322	0
	bra.uni 	$Lt_48_2495746;
$Lt_48_2496002:
	.loc	20	490	0
	mul.ftz.f32 	%f176, %f22, %f19;
	sub.ftz.f32 	%f177, %f1, %f176;
	mul.ftz.f32 	%f178, %f22, %f18;
	sub.ftz.f32 	%f179, %f127, %f178;
	mul.ftz.f32 	%f180, %f22, %f17;
	sub.ftz.f32 	%f181, %f3, %f180;
	mul.ftz.f32 	%f182, %f177, %f16;
	fma.rn.ftz.f32 	%f183, %f15, %f179, %f182;
	fma.rn.ftz.f32 	%f184, %f14, %f181, %f183;
	mul.ftz.f32 	%f185, %f177, %f13;
	fma.rn.ftz.f32 	%f186, %f12, %f179, %f185;
	fma.rn.ftz.f32 	%f187, %f11, %f181, %f186;
	mul.ftz.f32 	%f188, %f177, %f10;
	fma.rn.ftz.f32 	%f189, %f9, %f179, %f188;
	fma.rn.ftz.f32 	%f190, %f8, %f181, %f189;
	.loc	20	505	0
	mul.ftz.f32 	%f191, %f187, %f40;
	.loc	20	506	0
	mul.ftz.f32 	%f192, %f190, %f40;
	.loc	20	507	0
	mul.ftz.f32 	%f193, %f184, %f40;
	.loc	22	322	0
	mov.f32 	%f146, %f193;
	mov.f32 	%f147, %f192;
	mov.f32 	%f148, %f191;
$Lt_48_2495746:
	mov.f32 	%f194, %f76;
	bra.uni 	$Lt_48_2479874;
$Lt_48_2480130:
	.loc	20	490	0
	mul.ftz.f32 	%f195, %f22, %f19;
	sub.ftz.f32 	%f196, %f47, %f195;
	mul.ftz.f32 	%f197, %f22, %f18;
	sub.ftz.f32 	%f198, %f78, %f197;
	mul.ftz.f32 	%f199, %f22, %f17;
	sub.ftz.f32 	%f200, %f44, %f199;
	mul.ftz.f32 	%f201, %f196, %f16;
	fma.rn.ftz.f32 	%f202, %f15, %f198, %f201;
	fma.rn.ftz.f32 	%f203, %f14, %f200, %f202;
	mul.ftz.f32 	%f204, %f196, %f13;
	fma.rn.ftz.f32 	%f205, %f12, %f198, %f204;
	fma.rn.ftz.f32 	%f206, %f11, %f200, %f205;
	mul.ftz.f32 	%f207, %f196, %f10;
	fma.rn.ftz.f32 	%f208, %f9, %f198, %f207;
	fma.rn.ftz.f32 	%f209, %f8, %f200, %f208;
	.loc	20	505	0
	mul.ftz.f32 	%f210, %f206, %f40;
	.loc	20	506	0
	mul.ftz.f32 	%f211, %f209, %f40;
	.loc	20	507	0
	mul.ftz.f32 	%f212, %f203, %f40;
	.loc	22	322	0
	mov.f32 	%f97, %f212;
	mov.f32 	%f98, %f211;
	mov.f32 	%f99, %f210;
$Lt_48_2479874:
	mov.f32 	%f213, %f76;
	bra.uni 	$Lt_48_2464002;
$Lt_48_2464258:
	.loc	20	490	0
	mov.f32 	%f214, 0f447fc000;   	// 1023
	mov.f32 	%f215, 0f437f0000;   	// 255
	div.approx.ftz.f32 	%f216, %f214, %f215;
	mul.ftz.f32 	%f217, %f216, %f19;
	sub.ftz.f32 	%f218, %f6, %f217;
	mul.ftz.f32 	%f219, %f216, %f18;
	sub.ftz.f32 	%f220, %f7, %f219;
	mul.ftz.f32 	%f221, %f216, %f17;
	sub.ftz.f32 	%f222, %f5, %f221;
	mul.ftz.f32 	%f223, %f218, %f16;
	fma.rn.ftz.f32 	%f224, %f15, %f220, %f223;
	fma.rn.ftz.f32 	%f225, %f14, %f222, %f224;
	mul.ftz.f32 	%f226, %f218, %f13;
	fma.rn.ftz.f32 	%f227, %f12, %f220, %f226;
	fma.rn.ftz.f32 	%f228, %f11, %f222, %f227;
	mul.ftz.f32 	%f229, %f218, %f10;
	fma.rn.ftz.f32 	%f230, %f9, %f220, %f229;
	fma.rn.ftz.f32 	%f231, %f8, %f222, %f230;
	.loc	20	505	0
	mov.f32 	%f232, 0f437f0000;   	// 255
	mov.f32 	%f233, 0f447fc000;   	// 1023
	div.approx.ftz.f32 	%f234, %f232, %f233;
	mul.ftz.f32 	%f41, %f234, %f228;
	.loc	20	506	0
	mul.ftz.f32 	%f42, %f234, %f231;
	.loc	20	507	0
	mul.ftz.f32 	%f43, %f234, %f225;
	.loc	22	322	0
	mov.f32 	%f235, 0f3f800000;   	// 1
	mov.f32 	%f236, 0f447fc000;   	// 1023
	div.approx.ftz.f32 	%f237, %f235, %f236;
	mov.f32 	%f238, 0f447fc000;   	// 1023
	mul.ftz.f32 	%f76, %f237, %f238;
$Lt_48_2464002:
	ld.param.s32 	%r67, [__cudaparm_PixelFormatConvert_IR_PixelFormat_V210_422_10u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDestDeviceFormat];
	mov.s32 	%r68, 0;
	setp.eq.s32 	%p6, %r67, %r68;
	ld.param.s32 	%r69, [__cudaparm_PixelFormatConvert_IR_PixelFormat_V210_422_10u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDestPitch];
	mul.lo.s32 	%r70, %r69, %r10;
	add.s32 	%r71, %r11, %r70;
	cvt.s64.s32 	%rd5, %r71;
	ld.param.u64 	%rd6, [__cudaparm_PixelFormatConvert_IR_PixelFormat_V210_422_10u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel_inDest];
	@!%p6 bra 	$Lt_48_2535426;
	.loc	19	126	0
	mul.lo.u64 	%rd7, %rd5, 8;
	add.u64 	%rd8, %rd6, %rd7;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f43;
	mov.b32		%r72, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f42;
	mov.b32		%r73, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f41;
	mov.b32		%r74, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f76;
	mov.b32		%r75, %b1; }
	st.global.v4.u16 	[%rd8+0], {%r72,%r73,%r74,%r75};
	.loc	22	322	0
	bra.uni 	$Lt_48_2535170;
$Lt_48_2535426:
	.loc	19	126	0
	mul.lo.u64 	%rd9, %rd5, 16;
	add.u64 	%rd10, %rd6, %rd9;
	st.global.v4.f32 	[%rd10+0], {%f43,%f42,%f41,%f76};
$Lt_48_2535170:
	.loc	22	322	0
	@!%p2 bra 	$Lt_48_2535682;
	mov.f32 	%f239, %f77;
	mov.f32 	%f240, %f71;
	mov.f32 	%f241, %f70;
	mov.f32 	%f242, %f69;
	@!%p6 bra 	$Lt_48_2536450;
	.loc	19	126	0
	mul.lo.u64 	%rd11, %rd5, 8;
	add.u64 	%rd8, %rd6, %rd11;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f242;
	mov.b32		%r76, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f241;
	mov.b32		%r77, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f240;
	mov.b32		%r78, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f239;
	mov.b32		%r79, %b1; }
	st.global.v4.u16 	[%rd8+8], {%r76,%r77,%r78,%r79};
	.loc	22	322	0
	bra.uni 	$Lt_48_2536194;
$Lt_48_2536450:
	.loc	19	126	0
	mul.lo.u64 	%rd12, %rd5, 16;
	add.u64 	%rd10, %rd6, %rd12;
	st.global.v4.f32 	[%rd10+16], {%f242,%f241,%f240,%f239};
$Lt_48_2536194:
	.loc	22	322	0
	add.s32 	%r80, %r11, 2;
	setp.le.s32 	%p7, %r12, %r80;
	@%p7 bra 	$Lt_48_2536706;
	mov.f32 	%f243, %f213;
	mov.f32 	%f244, %f99;
	mov.f32 	%f245, %f98;
	mov.f32 	%f246, %f97;
	@!%p6 bra 	$Lt_48_2537474;
	.loc	19	126	0
	mul.lo.u64 	%rd13, %rd5, 8;
	add.u64 	%rd8, %rd6, %rd13;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f246;
	mov.b32		%r81, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f245;
	mov.b32		%r82, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f244;
	mov.b32		%r83, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f243;
	mov.b32		%r84, %b1; }
	st.global.v4.u16 	[%rd8+16], {%r81,%r82,%r83,%r84};
	.loc	22	322	0
	bra.uni 	$Lt_48_2537218;
$Lt_48_2537474:
	.loc	19	126	0
	mul.lo.u64 	%rd14, %rd5, 16;
	add.u64 	%rd10, %rd6, %rd14;
	st.global.v4.f32 	[%rd10+32], {%f246,%f245,%f244,%f243};
$Lt_48_2537218:
	.loc	22	322	0
	add.s32 	%r85, %r11, 3;
	setp.le.s32 	%p8, %r12, %r85;
	@%p8 bra 	$Lt_48_2537730;
	mov.f32 	%f247, %f126;
	mov.f32 	%f248, %f125;
	mov.f32 	%f249, %f124;
	mov.f32 	%f250, %f123;
	@!%p6 bra 	$Lt_48_2538498;
	.loc	19	126	0
	mul.lo.u64 	%rd15, %rd5, 8;
	add.u64 	%rd8, %rd6, %rd15;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f250;
	mov.b32		%r86, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f249;
	mov.b32		%r87, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f248;
	mov.b32		%r88, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f247;
	mov.b32		%r89, %b1; }
	st.global.v4.u16 	[%rd8+24], {%r86,%r87,%r88,%r89};
	.loc	22	322	0
	bra.uni 	$Lt_48_2538242;
$Lt_48_2538498:
	.loc	19	126	0
	mul.lo.u64 	%rd16, %rd5, 16;
	add.u64 	%rd10, %rd6, %rd16;
	st.global.v4.f32 	[%rd10+48], {%f250,%f249,%f248,%f247};
$Lt_48_2538242:
	.loc	22	322	0
	add.s32 	%r90, %r11, 4;
	setp.le.s32 	%p9, %r12, %r90;
	@%p9 bra 	$Lt_48_2538754;
	mov.f32 	%f251, %f194;
	mov.f32 	%f252, %f148;
	mov.f32 	%f253, %f147;
	mov.f32 	%f254, %f146;
	@!%p6 bra 	$Lt_48_2539522;
	.loc	19	126	0
	mul.lo.u64 	%rd17, %rd5, 8;
	add.u64 	%rd8, %rd6, %rd17;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f254;
	mov.b32		%r91, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f253;
	mov.b32		%r92, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f252;
	mov.b32		%r93, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f251;
	mov.b32		%r94, %b1; }
	st.global.v4.u16 	[%rd8+32], {%r91,%r92,%r93,%r94};
	.loc	22	322	0
	bra.uni 	$Lt_48_2539266;
$Lt_48_2539522:
	.loc	19	126	0
	mul.lo.u64 	%rd18, %rd5, 16;
	add.u64 	%rd10, %rd6, %rd18;
	st.global.v4.f32 	[%rd10+64], {%f254,%f253,%f252,%f251};
$Lt_48_2539266:
	.loc	22	322	0
	add.s32 	%r95, %r11, 5;
	setp.le.s32 	%p10, %r12, %r95;
	@%p10 bra 	$Lt_48_2539778;
	mov.f32 	%f255, %f175;
	mov.f32 	%f256, %f174;
	mov.f32 	%f257, %f173;
	mov.f32 	%f258, %f172;
	@!%p6 bra 	$Lt_48_2540546;
	.loc	19	126	0
	mul.lo.u64 	%rd19, %rd5, 8;
	add.u64 	%rd8, %rd6, %rd19;
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f258;
	mov.b32		%r96, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f257;
	mov.b32		%r97, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f256;
	mov.b32		%r98, %b1; }
	{ .reg .b32 %b1;
	cvt.rn.ftz.f16.f32	%b1, %f255;
	mov.b32		%r99, %b1; }
	st.global.v4.u16 	[%rd8+40], {%r96,%r97,%r98,%r99};
	.loc	22	322	0
	bra.uni 	$Lt_48_2540290;
$Lt_48_2540546:
	.loc	19	126	0
	mul.lo.u64 	%rd20, %rd5, 16;
	add.u64 	%rd10, %rd6, %rd20;
	st.global.v4.f32 	[%rd10+80], {%f258,%f257,%f256,%f255};
$Lt_48_2540290:
$Lt_48_2539778:
$Lt_48_2538754:
$Lt_48_2537730:
$Lt_48_2536706:
$Lt_48_2535682:
$Lt_48_2463490:
	.loc	22	322	0
	exit;
$LDWend_PixelFormatConvert_IR_PixelFormat_V210_422_10u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel:
	} // PixelFormatConvert_IR_PixelFormat_V210_422_10u_709_To_IR_PixelFormat_BGRA_4444_32f_Kernel
	.const .align 4 .b8 kRGB32f_To_YIQ[36] = {135,22,153,62,162,69,22,63,213,120,233,61,216,128,24,63,27,133,140,190,149,124,164,190,236,135,88,62,134,200,5,191,22,77,159,62};
	.const .align 4 .b8 kYIQ_To_RGB32f[36] = {0,0,128,63,20,208,116,63,219,249,30,63,0,0,128,63,177,80,139,190,2,188,37,191,0,0,128,63,45,178,141,191,85,48,218,63};

